Metaglossia: The Translation World

A study of 100,000 people shows AI can match average creativity—but the most creative humans still outperform machines.AI Can Match Average Human Creativity—But We Still Hold the Edge Where It Matters Most, New Study Finds
Tim McMillan·January 27, 2026
Advances in artificial intelligence have fueled a growing belief that machines are on the verge of matching, or even surpassing, human creativity. Large language models can now write poems, spin short stories, and generate clever wordplay in seconds. To many, these outputs feel creative enough to blur the line between human imagination and machine-generated language.

However, a new large-scale empirical study suggests that while today’s most advanced AI systems can rival the average human on certain creativity measures, they still fall short of the most creative minds—and that gap remains significant.

 

The research, published in Scientific Reports, offers one of the most comprehensive head-to-head comparisons yet between human creativity and large language models (LLMs).

By benchmarking multiple AI systems against a dataset of 100,000 human participants, the study moves the conversation beyond anecdotes and viral examples, replacing speculation with quantitative evidence.

“Our study shows that some AI systems based on large language models can now outperform average human creativity on well-defined tasks,” co-author and Professor at the University of Montreal, Dr. Karim Jerbi, said in a press release. “This result may be surprising — even unsettling — but our study also highlights an equally important observation: even the best AI systems still fall short of the levels reached by the most creative humans.”

Measuring creativity without subjectivity
One of the biggest challenges in studying creativity is measurement. Traditional creativity tests often rely on subjective ratings or expert judgment, making comparisons between humans and machines difficult.

To address this, researchers focused on divergent thinking—the ability to generate ideas meaningfully different from one another—using computational tools that can be consistently applied to both people and AI.

Researchers focused on the Divergent Association Task (DAT), a well-established test in creativity research. Participants are asked to generate a list of words that are as different in meaning as possible.

The creativity score is then calculated using semantic distance: how far apart the words are from each other in a high-dimensional language space. Larger distances indicate more diverse associations and, by extension, higher divergent creativity.

“These divergence-based measures index associative thinking—the ability to access and combine remote concepts in semantic space—an established facet of creative cognition,” researchers explain. 

Because the DAT relies on objective mathematical scoring rather than human judgment, it offers a rare opportunity to compare human and machine creativity on equal footing.

Where AI beats the average human
Using identical scoring methods, researchers tested a wide range of language models, including GPT-4, GPT-4-Turbo, Gemini Pro, Claude, and several open-source systems. The results were intriguing.

On average, some models—most notably GPT-4—matched or even exceeded the mean DAT score of the human population. Gemini Pro performed statistically on par with humans, while several other models fell just below.

This finding helps explain why AI-generated text can feel surprisingly inventive. Compared with everyday human performance, top-tier models can produce word combinations spanning a wide semantic territory. In purely statistical terms, they can appear impressively creative.

However, the study emphasizes that this is only part of the story.

The ceiling AI still cannot break
When the researchers ranked human participants by their performance on the Divergent Association Task and compared AI systems against those highest-scoring groups, a clear gap emerged. Humans in the top 50% of DAT scores, and especially those in the top 10%, consistently outperformed every tested language model.

Even GPT-4, the highest-performing system in the study, failed to reach the creativity levels of humans in the upper tail of the distribution.

“The top performing LLMs are still largely surpassed by the aggregated top half of human participants, underscoring a ceiling that current LLMs still fail to surpass,” researchers write.

In practical terms, this suggests that while AI can imitate creative behaviors common in the general population, it struggles to replicate the kind of associative leaps made by highly creative individuals—writers, artists, and innovators who regularly push conceptual boundaries.

Creativity can be tuned—but only so far
The researchers also explored whether AI creativity could be enhanced through technical adjustments. By increasing a model’s “temperature”—a parameter that controls randomness in word selection—they observed higher creativity scores. Higher temperatures reduced repetitive word choices and encouraged broader exploration of semantic space.

Prompting strategies mattered as well. When models were explicitly instructed to use certain approaches, such as drawing on word etymology, their divergent creativity scores improved. These findings suggest that AI creativity is not fixed, but responsive to how models are guided and configured.

Still, these gains did not eliminate the gap between AI and the most creative humans. Tuning improved performance, but it did not fundamentally change the ceiling.

From word lists to stories and poems
To test whether DAT performance translated into more realistic creative tasks, the study also examined creative writing. Models were asked to generate haikus, movie synopses, and flash fiction, which were then evaluated using automated measures of semantic diversity and textual complexity.

Here again, AI systems demonstrated impressive abilities, particularly in longer formats such as synopses and short stories. Yet human-written texts retained a consistent advantage, especially in structured, stylistically constrained forms such as haiku.

The results suggest that while AI can approximate certain surface features of creative writing, deeper patterns of originality and integration remain more characteristic of human authors.

What this means for creative work
The findings cut against both extremes of the AI creativity debate. On the one hand, they undermine claims that creativity is an exclusively human trait beyond the reach of machines. On the other hand, they challenge fears that AI is poised to replace top-tier creative professionals.

“The persistent gap between the best-performing humans and even the most advanced LLMs indicates that the most demanding creative roles in industry are unlikely to be supplanted by current artificial intelligence systems,” researchers write.

Rather than a story of replacement, the research points toward collaboration. AI systems may serve as creative amplifiers—useful for brainstorming, variation, and exploration—while humans remain essential for the highest levels of originality, judgment, and synthesis.

A clearer picture of human creativity
Beyond AI benchmarking, the study also reframes how scientists think about creativity itself. If machines can score well on some creativity tests without human-like cognition, it raises important questions about what those tests truly measure.

By combining large human datasets with objective semantic metrics, the research offers a more nuanced view: creativity is not a single ability, but a spectrum. AI may operate convincingly in the middle of that spectrum, while the upper extremes remain distinctly human. At least for now.

As AI systems continue to evolve, this kind of rigorous, data-driven comparison will be essential. Ultimately, the evidence is clear that machines can imitate creativity, but the deepest wells of human imagination are not so easily replicated.

“Even though AI can now reach human-level creativity on certain tests, we need to move beyond this misleading sense of competition,” Dr. Jerbi says. “Generative AI has above all become an extremely powerful tool in the service of human creativity: it will not replace creators, but profoundly transform how they imagine, explore, and create — for those who choose to use it.”

Tim McMillan is a retired law enforcement executive, investigative reporter and co-founder of The Debrief. His writing typically focuses on defense, national security, the Intelligence Community and topics related to psychology. You can follow Tim on Twitter: @LtTimMcMillan.  Tim can be reached by email: tim@thedebrief.org or through encrypted email: LtTimMcMillan@protonmail.com "
https://thedebrief.org/ai-can-match-average-human-creativity-but-we-still-hold-the-edge-where-it-matters-most-new-study-finds/
#Metaglossia
#metaglossia_mundus
#métaglossie

"Translator and interpreter Elsa Pérez and biologist Justin Jones provide expertise for Spanish-language 'Camaleon' podcast, teaching thousands of budding science enthusiasts around the world about a mysterious marine creature.

By Mary-Ann Muffoletto | January 27, 2026

If you were recruiting an ambassador to get kids excited about science, you might not immediately think of the homely hagfish. Eel-shaped with tentacles and a protruding mouth, the primitive marine creature, which emits gooey slime to foil predators, presents a menacing appearance.

And yet, the unusual sea critter is wowing youngsters in the episode “Pez Bruja (Hagfish)” on Spotify and YouTube, through the Spanish-language podcast Camaleón: Animales para niños curiosas (Chameleon: Animals for Curious Children). The creative effort features the volunteered expertise of Utah State University faculty members Elsa Pérez and Justin Jones.

Geared to children ages 3 to 10, Camaléon is a series of 20-minute podcasts about varied animals produced by Austin, Texas-based production company Cumbre Media and hosted by podcaster Gerwuin Riera. Each episode focuses on questions submitted by children and answered by zoo and university experts. The Pez Bruja episode, featuring Pérez and Jones, was released in November 2025 and has garnered more than 8,700 plays, representing more than 3,000 hours of STEM engagement, along with 20,000 hours of listening by children and parents in multiple countries.

“This is great news — we’re glad our episode is resonating with listeners,” says Pérez, assistant professor in the Department of World Languages and Cultures and co-director of the translation and interpretation studies degree program. “The producers told us it has been one of their most popular episodes.”

Pérez teamed with Jones, associate professor in the Department of Biology, at the suggestion of the USU Office of Research, which was approached by Cumbre Media.

Jones, who is director of USU’s Spider Silk Laboratory on Innovation Campus, studies hagfish intermediate filament proteins for their potential in creating novel biomaterials for varied applications in medicine, manufacturing, materials science and defense.

“I was excited by the chance to share real science through storytelling in a way that’s accessible and engaging for kids,” Jones says. “The hagfish is such a strange and fascinating organism that it naturally sparks curiosity, making it an ideal gateway into biology beyond the classroom.”

That gateway originated from actual questions from children in Mexico, Costa Rica, Spain, the United States and other countries, submitted to Camaléon producers. Jones wrote responses, which he shared with Pérez for translation and interpretation.

The process, Pérez says, meant not only learning about the unique characteristics of hagfish from Jones and translating them to Spanish, but expressing the information is a way that was engaging and accessible for young children.

“I carefully chose words children would understand and use themselves,” she says. “The podcast is driven by the children’s questions, so it was important to focus on the knowledge they were seeking.”

The name of creature itself, Pez Bruja, evoked some trepidation about the animal.

“Bruja means ‘witch’ in Spanish and several children described the creature as ‘scary, magical and ugly’ in their questions,” Pérez says. “But that was an opportunity to explore the animal and dispel some myths.”

For example, the children learned that by feeding on dead marine creatures — which sounds creepy, at first — hagfish perform a vital ecological service by functioning, essentially, as oceanic garbage disposals.

“From there, I described how hagfish are being studied in Justin’s lab to produce products that will help people and make their lives better,” Pérez says. “The process of making scientific knowledge fun and exciting for children was very rewarding.”

Jones applauds the opportunity to communicate science in more than one language.

“Science should be accessible to everyone, and language should never be a barrier to curiosity and learning,” he says. “It was genuinely fun and refreshing to step outside the traditional academic space. Working with Elsa and the Camaléon team was a great experience. I was surprised how creative the process was and how effective storytelling can be for communicating science.”

Jones says that working with young children is especially meaningful to him.

“Younger kids are still forming how they view science,” he says. “At that stage, small moments can have a big impact. Engaging them early gives us the chance to plant seeds of curiosity that can grow over time.”

Pérez says she was inspired by the multidisciplinary effort.

“It opened my eyes to all the amazing knowledge, research and talent we have within the College of Arts and Sciences,” she says. “It showed me what amazing things we can accomplish when we unite in a common goal to serve the greater community.”

WRITER
Mary-Ann Muffoletto
Communications Specialist
College of Arts & Sciences
435-797-3517
maryann.muffoletto@usu.edu

CONTACT
Elsa Pérez
Assistant Professor of Translation and Interpretation
College of Humanities and Social Sciences
801-358-9085
elsa.perez@usu.edu

Justin Jones
Associate Professor
Department of Biology
435-797-9292
justin.a.jones@usu.edu"
https://www.usu.edu/today/story/curious-kids-usu-interpreter-and-scientist-team-to-create-science-learning-magic-for-children
#Metaglossia
#metaglossia_mundus
#métaglossie

"...Vous écrivez en langue amazighe. Que représente cette langue pour vous en tant que jeune romancière ?
J’écris en langue amazighe, plus précisément en kabyle, parce que c’est ma langue maternelle et qu’elle mérite d’être vivante et entendue. Je vois combien d’autres langues ont leurs écrivains, leurs institutions, leur richesse littéraire. Mon choix n’est pas par ignorance des autres langues, mais parce que notre langue a besoin de nous : elle a besoin qu’on la fasse vivre, qu’on l’aide à avancer et à libérer des idées qui y sont encore trop souvent prisonnières.

Pensez-vous que la littérature amazighe est aujourd’hui en train de se renouveler ?
Bien sûr. Avec ses chercheurs, ses écrivains et ses traducteurs, la littérature amazighe se renouvelle et s’ouvre à de nouvelles formes et horizons. Elle a aujourd’hui le potentiel d’aller bien au-delà de ses frontières, de toucher un public universel et de faire entendre sa richesse partout dans le monde.

Quel message souhaitez-vous laisser au lecteur à travers ce roman ?
À travers Les fragments de la mémoire et Les chambres du cerveau, je souhaite inviter le lecteur à réfléchir sur la conscience et l’identité. La mort n’y est pas une fin, mais une traversée, un passage où l’identité se défait pour laisser émerger une autre forme de conscience. C’est avant tout une exploration de ce que signifie être vivant et penser.

Après ce prix, comment envisagez-vous la suite de votre parcours, entre sciences et littérature ?
Ce prix m’encourage surtout à continuer d’écrire. J’ai beaucoup d’envies : explorer de nouveaux thèmes, publier d’autres romans et approfondir mes réflexions sur le cerveau, la mémoire et la conscience. Je souhaite poursuivre ce lien entre sciences et littérature, en laissant ces deux univers se nourrir mutuellement dans mes prochains ouvrages. Y. Z."

https://elwatan.dz/lisa-serik-laureate-du-prix-du-president-de-la-litterature-et-la-langue-amazighes-recevoir-un-tel-prix-est-un-encouragementa-continuer-doser/
#Metaglossia
#metaglossia_mundus
#métaglossie

"Minister of State for European Affairs and Defence, Thomas Byrne TD, has today (26 January) signed an Administrative Cooperation Agreement, for the provision of simultaneous interpretation services, including Irish, at informal Ministerial meetings and summits to be held in Ireland during the EU Presidency.

 

The agreement marks a significant milestone in Ireland’s linguistic and cultural engagement with the European Union and provides an opportunity to reaffirm Ireland’s commitment to multilingualism, and the full integration of the Irish language, in European affairs.

 

Minister Byrne was joined by the Director-General of the European Commission’s Directorate-General for Interpretation, Ms Genoveva Ruiz Calavera, for the signing of the agreement.

 

Speaking at the signing, Minister Byrne said: “Ireland’s 2026 EU Presidency is a landmark because it is our first EU Presidency since the Irish language assumed its full status as an official and working language of the EU. As a result of this agreement, Irish will be used at informal EU Presidency Ministerial meetings and summits held in Ireland.”

 

Minister Byrne added: “The signing of this agreement reflects our ongoing and intense preparations for Ireland’s forthcoming Presidency of the Council of the European Union, providing the linguistic framework necessary for a successful term. I believe that the ability for Ministers and officials, from across the EU, to engage fully in their own languages supports meaningful dialogue, effective negotiations and good decision making.”

 

The Directorate‑General for Interpretation team of interpreters has extensive experience of providing conference interpretation in the 24 official languages of the European Union, and of working with successive rotating Presidencies of the EU Council.

 

The capacity of the Directorate-General for Interpretation to provide simultaneous interpretation services in the Irish language during the upcoming Presidency will support these efforts.

26 January 2026

 

Notes

DG SCIC is the European Commission’s interpreting service. It is the world’s largest interpreting service, routinely providing interpretation in all 24 official EU languages for the European Commission, European Council, Council of the EU, rotating Presidencies of the Council of the EU, and other EU bodies and agencies.

 

Ireland last held the Presidency of the Council of the EU from January to June 2013, at a time when Irish was an official and working language of the EU but still subject to a derogation limiting the scope of translation and interpretation services. This derogation was lifted on 1 January 2022."

https://www.gov.ie/en/department-of-foreign-affairs/press-releases/minister-byrne-signs-interpretation-agreement-for-irelands-eu-presidency/

#Metaglossia 

#metaglossia_mundus 

#métaglossie 

"The UB launches a pioneering vocabulary of scientific terminology in Catalan Sign Language (LSC)

 

A deaf student can follow university classes with a sign language interpreter, but they often encounter the problem that there are no specific signs for many of the scientific terms used in the classroom. In these cases, the word must be spelled out or a sign must be “agreed upon”, even if only for that particular occasion.

 

The University of Barcelona has launched a pioneering project: a vocabulary of scientific terms in Catalan Sign Language (LSC). The aim is to make university classes more accessible to deaf signers and to promote equal opportunities in the academic sphere. This vocabulary is published as videos on the UBTERM terminology portal, managed by the UB Language Services, which have also created a web portal of resources on LSC.

 

“For each term, there must be a clear and easy-to-understand sign that also respects the phonological structure of Catalan Sign Language,” SAYS Raúl Ruiz, coordinator of the project’s team, which includes professors, interpreters, and linguists specialized in LSC. This Biochemistry student and Catalan Sign Language professor emphasizes that, until now, there has been no structured initiative to develop specialized scientific terminology in LSC at the university level, beyond occasional glossaries or specific thematic projects. This contrasts with the existence, for years, of well-established resources in other sign languages, such as American (ASL) or British (BSL).

 

A powerful tool for understanding science

 

To carry out this task, a process had to be developed from scratch: first, the term in Catalan is selected, and a search is conducted among existing LSC signs to see if an equivalent already exists or if a new sign needs to be created, based on the resources of Catalan Sign Language. At the same time, the concept is analyzed with the help of professors who are experts in the specific discipline to which the term belongs. Once this work is completed, the sign is created if necessary and validated by experts and users from scientific fields in LSC. Finally, the recording is made and reviewed again by experts. Once the accuracy of the signed term recording is confirmed, it is published on the UBTERM portal, included as an additional equivalent in the corresponding terminological entry.

 

Currently, the first published signed terms — 59 in total, all from the field of mathematics — can already be consulted. For each term, there is a GIF file and three videos showing the same signed term from three different angles: front, from the right, and from the left, to facilitate visual perception.

 

A web portal WITH resources in LSC is also created to facilitate learning and promote equal opportunities.

After receiving support from the Vice-Rector’s Offices for Teaching Policy and Language Policy, the Language Services have taken on the management and publication of the project. They have done so as part of their regular terminological work, but with the particularity that this time a signed language is involved, which requires a specific approach. Àngels Egea, head of Terminology at the Language Services, recalls: “As with any other language, updating the terminological repertoire is an ongoing task with no end date. This means that this newly launched project does not have a fixed timeline, but rather a long road ahead.”

 

Raúl Ruiz emphasizes that the project is important not only for the deaf signing community but also for society as a whole: “Catalan Sign Language can become a very powerful tool for understanding scientific concepts, thanks to its visual nature and use of space. For example, in fields such as mathematics, where we work with graphs, coordinates, or X, Y, and Z axes, the fact that LSC uses three-dimensional space greatly facilitates the representation of these concepts. In this sense, LSC not only makes science more accessible to deaf signers, but also enriches the way anyone can understand and communicate scientific knowledge.”

 

This strategic project is part of the initiatives funded by the Government of Catalonia’s Interlingua grants, which aim to promote and support the use of languages in the inter-university context."

(27/01/2026)

https://web.ub.edu/en/web/actualitat/w/scientific-terminology-catalan-sign-language

#Metaglossia 

#metaglossia_mundus 

#métaglossie 

"L’auteur et éditeur Frédéric Boyer signe une nouvelle traduction d'Hamlet, pour la mise en scène qu’en propose Ivo Van Hove. L’occasion de relire avec lui la plus fameuse des tragédies de Shakespeare et de faire entendre sa portée contemporaine.

Avec
Frédéric Boyer, écrivain, traducteur et éditeur
Dans La Tragédie d'Hamlet, Prince du Danemark cette pièce parue au tout début du XVIIe siècle, le prince du Danemark a tout récemment perdu son père, il est chamboulé quand il apprend que son spectre rôde. Le fantôme révèle bientôt la véritable cause de sa mort, il a été assassiné par son propre frère Claudius, désormais roi à la place du roi, et remarier à la reine, sa veuve. Hamlet veut venger la mémoire de son père, il singe la folie, délaisse son amoureuse et imagine une pièce de théâtre piège, susceptible de forcer les uns les autres à la vérité. Ce texte est une tragédie sur le langage, l'action et la lâcheté.

Une pièce baroque
Frédéric Boyer explique que "ce qui fait d'Hamlet une pièce baroque est le rapport au temps que cette pièce propose. Il vrille, sort de son axe, est désarticulé. Hamlet est d'une prétention folle en se donnant pour mission de le remettre d'aplomb. Son chagrin et son deuil impossible le poussent à se donner cette grande responsabilité mais il n'y arrive pas du tout."

Le théâtre shakespearien, lieu de révélation
Hamlet est metteur en scène et dramaturge, au milieu de la propre pièce de Shakespeare au tout début du XVIIe. "À un moment il dit laissez entrer les comédiens, ils sont la quintessence de notre époque. Shakespeare fait alors du théâtre le lieu de la révélation."

Traduire une langue carnavalesque
Pour sa traduction de La Tragédie d'Hamlet, Prince du Danemark Frédéric Boyer essaye "de rester au plus proche du texte. Nous croyons que parce que nous traduisons Shakespeare, nous devons forcément trouver un langage qui soit à la fois châtié, académique et recherché. Finalement la langue shakespearienne est carnavalesque car elle mélange les registres et les niveaux de langue. À trop vouloir gommer cette langue dans la traduction, nous perdons quelque chose de très fort dans le texte shakespearien. C'est une affaire de choix."

La question de l'auditeur
La question d'Édouard @endoudlettres à l'attention de Frédéric Boyer  : "Est-ce que le choix que vous avez fait pour votre traduction d'une forme de nudité de la langue est volontaire pour rester au plus près de l'expérience intérieure qui se joue dans le personnage d'Hamlet, quitte un tout petit peu à perdre du lyrisme de la langue de Shakespeare ?""
https://www.radiofrance.fr/franceculture/podcasts/le-book-club/to-be-or-not-to-be-hamlet-retraduit-avec-frederic-boyer-4689542
#Metaglossia
#metaglossia_mundus
#métaglossie

"Multicultural identity in intercultural dialogue

The first event in the international online conference series entitled “Global Encounters: Intercultural Dialogues” was hosted by the Faculty of Humanities and Social Sciences and the Language and Communication Research Group at the University of Miskolc on January 22 and 23. The central themes of the discourse were intercultural communication, competence development, and identity construction.

 

BTK Campus Partners Research Science University

2026. January 26.

The conference was organized in cooperation with Chouaib Doukkali University (Morocco), Sunway University (Malaysia), St. Cyril and Methodius University of Veliko Tarnovo (Bulgaria), and II. Ferenc Rákóczi Hungarian University of Transcarpathia (Ukraine).

 

The two plenary lectures were given by Dr. Irina Golubeva (University of Baltimore, USA) and Dr. Tamás Kiss (Sunway University, Malaysia), who discussed the development of intercultural citizenship in the service of inclusive higher education and the encouragement of critical thinking through student questions.

 

A total of 33 presentations in eight sections examined issues of inter- and multicultural identity construction from the perspectives of literary representation, translation and the responsibility of translators, as well as professional and educational contexts.

 

Sixty participants from eight countries shared their research findings and educational experiences and took part in inspiring professional discussions. The conference provided an excellent opportunity to strengthen cooperation between the participating universities, as a result of which the partners are planning to publish a joint conference volume and continue the conference series in 2027."

https://www.uni-miskolc.hu/en/news/multicultural-identity-in-intercultural-dialogue/

#Metaglossia 

#metaglossia_mundus 

#métaglossie 

"Artists Farida and Rasilaben weave together during the workshop, exchanging stitches shaped by generations.

In a sunlit wooden hall in New Delhi, women sit together with needles moving steadily through cloth. Some work while seated on sofas; others gather in small circles on the floor, stitching side by side. Children move freely through space — watching, laughing and forming friendships of their own. The workshop feels less like a training room and more like a shared home, filled with creativity, care, and conversation.

Artisans from Kutch, a district in Gujarat in Northwestern India, and Afghanistan have come together to exchange embroidery techniques passed down through generations. Mirrored Kutch motifs meet delicate Afghan traditional patterns as hands demonstrate, repeat and adapt each stitch. What begins as shared practice slowly grows into collaboration — and then into trust and friendship.

For Parween Kakkar, a 47-year-old Afghan refugee who arrived in India in 2017, the exchange offered more than technical learning. Working alongside Kutch artisans felt natural to her. “It is a cultural exchange,” she says. With support from UNHCR, Parween was able to display and sell her embroidered products, earning income while strengthening her confidence. Reflecting on the experience, she adds, “I am happy that two countries are coming together and learning from each other. In these 10 days, I forgot all the troubles of my life.”

Across the hall, Dharmila Ahir, a 32-year-old artisan leader from Kotay village in Kutch, guides younger women as they work. Embroidery has been part of her life since childhood — from stitching her own bridal attire to now leading a collective of 80 women. “We learned from each other — teaching, learning and creating together,” she says. Through shared learning, new market opportunities are beginning to emerge.

pix: Dharmila Ahir and Khalida Mahmodi co-create a joint embroidery piece, blending Kutch and Afghan traditions.
© UNHCR / Paridhi Jain

Recalling how the collaboration began, Dharmila adds, “I learned about UNHCR and IICD’s workshop through the Kaarigar Clinic (an NGO in Gujarat helping women in marketing their products) and we were invited to work together — Kutch women and Afghan women. Even though we speak different languages, our hands speak the same language. In just 10 days, we learned from each other, created joint work and became good friends.”

Khalida Mahmodi also practiced Kutch embroidery techniques taught to her by Sitaben Rabari. For 20-year-old Rasila Rabari, the experience has been transformative. “I earn daily now,” she says. “I also found friendships that will stay with me.”

The collaboration has continued well beyond the workshop. Women from both communities stay connected through a shared WhatsApp group, where they exchange ideas, designs and client requests. Dharmila began promoting Afghan-made products on her Instagram page, introducing them to her existing customer base. When clients requested pocket purses, the message was shared across the group.

Parween and other Afghan women responded by creating the products collaboratively. As a result, Parween sold 10 pocket purses to a single client, earning an income and building confidence as an artisan entrepreneur. With more orders now coming in, the women are seeing stronger and more sustainable livelihood opportunities — created together across borders.

Taking the project forward

The workshop was part of Embroidery Without Borders, a collaborative initiative supported by UNHCR India and the Indian Institute of Crafts and Design (IICD), Jaipur. Over 10 days, Afghan refugee women and Kutch artisans exchanged embroidery skills and co-created contemporary, market-ready textile products. Alongside skill development, the initiative fostered dialogue, strengthened community connections, and supported pathways to sustainable livelihoods.

As the days passed, shared techniques transformed into embroidered pieces carrying stories of resilience, displacement, and belonging. The initiative shows how, when women from refugee and host communities create together, stitches can do more than decorate fabric — they can connect cultures and build shared futures with dignity.

Building on the success of this first phase, UNHCR India is seeking financial and additional support for the second phase of the project, with a focus on developing specific, high-quality products for wider markets. Discussions are also underway with MADE51 to explore pathways to global markets, while collaboration with IICD will continue to strengthen design, production and market readiness. to explore pathways to global markets, while collaboration with IICD will continue to strengthen design, production and market readiness.

MADE51 is a flagship initiative of UNHCR that empowers refugee artisans by promoting their skills, cultural heritage, and livelihoods through dignified work. UNHCR India has implemented MADE51 since 2018, working with local social enterprises to support more than 350 refugee artisans across three states.

With these efforts, the initiative is set to expand further in 2026 — creating more opportunities for refugee and host-community artisans alike."
https://www.unhcr.org/asia/news/stories/when-stitches-speak-across-borders-afghan-refugee-india
#Metaglossia
#metaglossia_mundus
#métaglossie

"Is AI Translating Tamil Literature? Publishers On New Technology

 

Sadek's book was translated into Tamil by publisher Senthil Nathan

A recent Tamil translation of Umberto Eco’s ‘The name of the rose’ made headlines not for its long-awaited release, but for criticism online that AI tools may have been used in the translation process.

 

Translator M D Muthukumaraswamy and publisher Anush Khan of Coimbatore-based Ethir Veliyeedu have rejected the allegation. “The translator has not used AI at any stage of the translation. Eco is known for his fragmented sentences, which the translator retained in Tamil. We did not expect that this would be seen as a sign of AI use,” says Anush. “Still, while we are not using AI for translations, we should not be wary of adopting new technologies.

 

”India has 9,000 publishers, but even among large publishing houses that release several hundred English-language titles annually, only a small fraction is translated into Indian languages, though there is a growing demand for regional-language content. Publishers say it is primarily because translation is time-consuming and resource-intensive. AI-powered solutions are now entering the scene to close that gap.

 

Bengaluru-based NAAV AI, for instance, launched in 2025 and uses models such as Claude and GPT-4 to support translations from English into Hindi, Marathi, Kannada, Tamil, Telugu, and Malayalam.

 

Chennai-based Ailaysa, a multilingual AI platform, has processed 200 books, reportedly charging about Rs 100 per 1,000 words for translation. It treats entire manuscripts as cohesive works instead of sentence-by-sentence conversions.

 

“We use AI to translate non-fiction and academic work as it saves cost and time,” says Senthil Nathan, CEO of Ailaysa and Aazhi Publishers. “For example, translators may need at least four to five months to translate a book. It also comes with hefty pay for the translator. With AI, the work is done in a week and with no cost. But we still need people to oversee the work.”

Senthil Nathan, who translated consumer psychologist Nadim Sadek’s book ‘Shimmer, Don’t Shake’ using his AI translation software, says he still had to edit the translation at the end. “It just bettered the end product. While AI works well for non-fiction, it’s still at a nascent stage for fiction or poetry. But it’s just a matter of time,” he says. “Another advantage of using AI tools is that you can translate Tamil work into English as well.”

At the recently concluded Chennai International Book Fair, Davide Daghia, director of digital at Bologna Children’s Book Fair, Italy, said publishers need to be cautious about using AI when it comes to children’s literature.

 

“Two years ago, Italian publisher Castoro got a story written by a children’s author and ChatGPT using the same prompt. While adults could tell the difference between the two stories, children could not. The AI-generated story followed all the rules of writing, but lacked the casual, illogical details that give children’s books their depth and appeal. AI cannot give you that human sensibility essential in writing for children.”

 

AI, though, can be helpful in marketing books, said Daghia. “In 2015, Italian author Michela Murgia created a Facebook page for the protagonist of her novel ‘Chiru’ to push sales. Chiru, the character, would write to followers and reply to their questions. I wrote to Chiru. Now imagine if you give this creative input to AI, a perfect communication plan can make your book sales explode,” said Daghia.

While Tamil publishing houses have just begun their AI journey, many foreign publishers are adding clauses to their contracts regarding AI use.

 

“Several publishers are cautious about using AI for content or design,” says Ananth Daksnamurthy, founder of Two Shores Press. “When they give translation rights for books, they insist AI should not be used. AI could still be useful in streamlining a publishing house’s workflow such as in copy editing and proofreading, and in correcting pronunciation errors in audiobooks.”

Writer Marudhan says AI does a better job of translating Tamil texts into English, not vice versa. “So now there is an opportunity where a Tamil writer with the help of AI can translate his own work into English,” he says. "But now comes the bigger question. Will the publishing industry say it is translated with the help of AI, or just use the author's or translator's name alone in the book covers?”"

Vinoth Kumar N / TNN / Jan 25, 2026

 

https://timesofindia.indiatimes.com/city/chennai/is-ai-translating-tamil-literature-publishers-on-new-technology/articleshow/127488965.cms

#Metaglossia 

#metaglossia_mundus 

#métaglossie 

Arunava Sinha recounts how his career in translation began with ‘Chowringhee’
"Celebrated translator Arunava Sinha on Sunday talked about his 100th published translation and the evolving craft of carrying Bengali literature to a global readership. In conversation with theatre director and cultural commentator Sujoy Prasad Chatterjee at the Kolkata Literary Meet, Sinha said much of his career began “almost by accident”.

Sinha traced the unlikely beginnings of his journey to Chowringhee. “I translated Chowringhee in 1992 — 30 years after the novel came out in Bangla,” he said. “And then it lay in cold storage for 14 years.” The manuscript resurfaced only when a Penguin editor asked the author Shankar for it. “He had forgotten who’d done it,” Sinha recalled. “Fortunately, I’d put my name on it. And that was how it all began”.

Asked about translating living authors, Sinha said the process rarely changes. “I translate any author the same way,” he said, whether they are alive or not. What differs, he said, is the occasional luxury of clarification — being able to call an author to ask what a word or phrase truly intends. That said, he added, such moments are rare.

The discussion turned to the difference between translating prose and poetry. “With poetry, it’s elliptical,” he said. “It’s probably a failed endeavour to try and arrive at a singular meaning.” A translator, he explained, must pay attention to “sound, rhythm, music, imagery,” often choosing to foreground only one or two elements. “That’s why two translators will never converge in poetry the way they sometimes do in prose.”

Asked why he chose translation over writing original fiction, Sinha said it wasn’t a conscious choice. “It’s not even a career that earns me a living,” he said, “It was just something I tried. I enjoyed it. Other people enjoyed it too.”

Timing, he added, mattered. “English-language publishing in India was just starting out. They were desperately looking for books — and not all of them could be written in English.”

Sinha acknowledged that a translator’s personal politics can clash with the worldview embedded in a text. Referring to a historical novel that portrays outsiders as enemies, he said, “My politics is opposed to the latent politics of the text.”

Still, he chose to translate it. “I didn’t want to take the easy option of not translating,” he said. Instead, he added a note: “Don’t shoot the messenger. The translator is only telling you what the text says in another language.”

‘All of India is in Kolkata’: Jhumpa Lahiri on language and the city that never left her..."
Agnivo Niyogi
Published 25.01.26
https://www.telegraphindia.com/my-kolkata/people/arunava-sinha-recounts-how-his-career-in-translation-began-with-chowringhee-at-kolkata-literary-meet/cid/2144324

"In an astonishing parallel to how human language evolves, communities in Mozambique use different ‘dialects’ to coordinate cooperation with wild birds.

In sub-Saharan Africa, some communities work together with wild birds to find honey. Using distinctive calls, honey-hunting people coordinate with greater honeyguides (Indicator indicator), who lead them to bee hives. A new study demonstrates that the honey-hunters’ calls vary between communities, and evolve in much the same way as human dialects.

Honey-hunters use two call types to communicate with honeyguides: recruitment calls to invite the wild birds to the hunt, and coordination calls to stay in contact with the bird as it leads them to a hive.

Can animals ask humans for help?
The weirdest animal friendships on the planet: Animal pairings that shouldn’t get along but are actually the best of friends
Once a hive has been located, honey-hunters use smoke to subdue the bees and harvest their honey, while the honeyguides help themselves to beeswax and tasty grubs.

Both species therefore benefit from this interaction: honeyguides get food that they couldn’t otherwise access, and honey-hunters are guided to the hive. Indeed, previous research shows that honey-hunters are over three times more likely to find a hive if they have a honeyguide guiding them.

Both types of call that the honey-hunters use – recruitment and coordination – vary across cultures, in the same way that human language varies. But even neighbouring villages have slightly different calls, a bit like regional dialects.

To find out if these ‘dialects’ evolved because of cultural processes, like human language dialects do, researchers recorded calls from 131 honey-hunters across 13 villages in Mozambique’s Niassa Special Reserve.

Can other species learn ‘foreign languages’?
Whales speak in dialects and elephants have names for each other: The incredible secrets of animal language
They found that neighbouring villages tended to have similar calls, while villages that were further apart had more dissimilar calls. Meanwhile, environmental factors had no effect on call similarity. This means that call variation is driven by cultural rather than environmental processes.

“The highlights just how powerfully culture shapes us as a species, even in our interactions with wild, untrained animals,” explains Dr Jessica Van Der Wal, lead author of the study.

While honey-hunters learn recruitment and coordination calls from their fathers, scientists are unsure how honeyguides learn to cooperate with humans.

“They can’t learn from their parents, because they are brood parasites: like cuckoos, they lay their eggs in other birds’ nests,” says Dr Van Der Wal. “But they may learn by observing other honeyguides interacting with humans.”

Senior author Professor Claire Spottiswoode, of the University of Cape Town’s FitzPatrick Institute of African Ornithology (who leads the Honeyguide Research Project), added “Humans learn and maintain the local signals needed to cooperate with honeyguides, and honeyguides are in turn probably learning and so helping to reinforce these local human dialects – much as they learn larger-scale variation in human signals across Africa, more akin to different human languages.”

Figuring out how honeyguides learn to cooperate with humans is one of the team’s next research questions."
Beki Hooper
January 26, 202
https://www.discoverwildlife.com/animal-facts/birds/greater-honeyguide-human-dialects
#Metaglossia
#metaglossia_mundus
#métaglossie

"Peter Burke once noted that translators, like historians, are “serving two masters and attempting to reconcile fidelity to the original with intelligibility to their readers” (2007). It is particularly interesting to examine how translation functioned during the emergence of a new political-economic reality and its language in the formation of a well-ordered police state. Among the texts on political economy translated from German into Russian, the most important were those on cameral and police sciences. The history of their translation is embedded in the broader history of cultural transfer (Espagne).

Manuscript translations of Wilhelm von Schröder (1640–1688) on the prince’s treasury from the early eighteenth century, the numerous multi-volume book and journal translations of Johann H. G. von Justi (1717–1771) on manufactures, welfare, and the science of state governance, and the influential translation of Joseph F. von Sonnenfels’s (1732–1817) book on politics and finance form the basis for reflections on the translation of the concepts of state, society, welfare, happiness (Glückseligkeit), and police (Policey). The paper is mainly focused on two questions: what was translated, and in what manner (with what intentions) were the translations carried out? At times, information about the translators and the intended audience provides additional context for understanding the adaptation, transfer, and reception of these ideas. In my research, I proceed from the non-neutrality of translation (not merely a linguistic transfer), but instead consider it as a process of profound transformation, appropriation, and semantic interaction. The political, aesthetic, and intellectual act of translation and transfer invites reflection on a broader “bracket” under the name of the Baroque that brings these translations together in the age of the Enlightenment.

Danila Raskov is a researcher at the Department of Economic History, Uppsala U... He is the author of the books "The Economic Institutions of Old Believers" (2012) and "The Rhetoric of Institutional Economics" (2023), both published in Russian. He is currently working on the book Cameralism in the Building of the Russian Empire: Administrative and Intellectual Discourses on the Well-Ordered State."
Date: 3 February 2026, 15:15–17:00
Location: IRES Library, Gamla torget 3, 3rd Floor
Type: Lecture, Seminar
Organiser: Institute for Russian and Eurasian Studies (IRES)
IRES higher seminar
Last modified: 2026-01-23

Contact
+46 18 471 00 00 (switchboard)
https://www.uu.se/en/department/russian-and-eurasian-studies/events/archive/2026-02-03-translations-of-cameralists-in-the-eighteenth-century-russian-empire
#Metaglossia
#metaglossia_mundus
#métaglossie

"March 3, 2026 | 6:00 - 7:00 pm
Enarson Classroom Building
Memes are more than jokes; they are cultural snapshots. This session explores how memes reflect humor, politics, identity, and social norms across different countries. Students will compare global meme trends and learn how culture shapes what goes viral and why.

The Office of International Affairs invites international and domestic students to join us for the weekly Global Engagement Night, where you can interact with one another and have cross-cultural discussions about different regions of the world and general topics affecting all college students.

All are welcome to join the conversation every Tuesday from 6-7 p.m.  in 160 Enarson Classroom Building."
https://oia.osu.edu/events/global-engagement-night-laughing-translation-global-meme-culture
#Metaglossia
#metaglossia_mundus
#métaglossie

"House Democrats Propose Bill to Mandate Multilingual Federal Resources

The bill is in direct response to the Trump administration’s efforts to make English the official language of the United States.

 

A group of Democratic House lawmakers want to require the federal government to offer information and written materials in multiple languages — a direct response to the Trump administration’s efforts to cut resources for non-English speakers.

 

Reps. Grace Meng, Judy Chu, Juan Vargas and Dan Goldman — members of the Congressional Asian Pacific American Caucus — on Friday will introduce legislation that would codify a Clinton-era executive order mandating that federal agencies offer resources to non-English speakers, according to bill text viewed by NOTUS.

 

“Every American deserves equal access to federal services and programs in a language they can

 

Understand,” Meng, who chairs the caucus, said in a statement. “We will continue to fight against the Trump administration’s attacks on immigrants and the essential services that our communities rely on and deserve.”

 

The effort comes after President Donald Trump’s March executive order designated English the official language of the United States. Trump’s directive rescinded the Clinton-era order and all accompanying policy guidance on how federal agencies could best serve non-English speakers.

 

The administration subsequently canceled some translation and interpretation contracts across multiple federal agencies.

 

Multiple federal resources for non-English speakers were discontinued or paused after Trump’s executive order. NOTUS reported in July that the Department of Justice “temporarily suspended” LEP.gov, the main site for non-English speakers to access multilingual materials including translated information from federal agencies. The site is still suspended more than six months later, though multilingual resources required by law are still available through other channels.

 

The Democrats’ bill would require federal agencies to ensure that “individuals with [limited English proficiency] can meaningfully access the Federally-conducted programs and activities,” including by translating documents into commonly spoken languages, offering interpretation services to non-English speakers and employing bilingual agency staff.

 

The legislation proposed Friday mandates that the federal government reopen access to LEP.gov.

 

The bill is unlikely to see widespread Republican support, with many GOP lawmakers cheering Trump’s executive order. But some Republicans who represent districts with high non-English-speaking populations said last year that they do not agree with Trump’s English-only efforts.

 

“I don’t understand why you’d make an issue out of it,” Rep. Carlos Gimenez, a Republican from Florida, told NOTUS in July. “I understand what the president’s trying to do with one language. But, if somebody feels more comfortable, or may speak English but reads Spanish, I don’t have a problem either way.”

 

The bill would also require the attorney general to create a system for people to submit complaints about barriers to language access in the federal government. LEP.gov previously aggregated links for each federal agency’s processes for filing discrimination complaints under the Civil Rights Act. Users can still submit those complaints, typically through an agency’s civil rights office, but there is no single portal or resource to do so.

 

Chu called the Trump administration’s actions “an attack on our immigrant communities.”

 

“In my district, translation services are essential for parents applying for a home loan, seniors accessing Medicare, immigrants starting a small business, and disaster survivors accessing the FEMA’s resources,” she said in a statement, adding that the proposed legislation would “ensure no one is denied health care, housing, or disaster assistance because English is not their first language.”"

Shifra Dayak

January 23, 2026

https://www.notus.org/congress/house-democrats-bill-mandate-multilingual-federal-resources

#Metaglossia 

#metaglossia_mundus 

#métaglossie 

"Red Cross and Capgemini implemented BabelSpeak to ensure that anyone with the need will be able to access translation services without worrying about the expense
Client challenge: In confidential conversations in which language is a barrier and that may involve sensitive, personal issues leave vulnerable groups – such as refugees – entirely dependent on interpreters to be understood.

Solution: Red Cross engaged Capgemini to build BabelSpeak, a tool for real-time translation and transcription in approximately 100 languages that simultaneously secures sensitive data and information.

Benefits:

Interpretation services made more accessible and efficient
Cost of interpretation reduced up to 90%
Every conversation is fully secured and kept private
Digital interpreter helps vulnerable individuals be understood in encounters with authorities
In collaboration with the Norwegian Red Cross, Capgemini has conducted extensive testing and improvement of the AI-based interpretation service BabelSpeak. By doing so, the partners have given a voice to marginalized groups in society and broken down linguistic barriers.

Endless need for translation
The need for interpretation services is vast. While the Norwegian government spends around NOK 1.4 billion annually on interpretation services, this is still not enough to meet demand.

The Red Cross in particular has an almost endless demand for the translation of sensitive conversations. These interpretation needs often involve deeply personal content – and in many cases, the conversations can be about sensitive issues that have major consequences for the people involved. However, as a non-profit organization, the Red Cross must work with a limited budget while professional translation often proves expensive.

This meant that it needed another option that came with a more affordable budget. To that end, the Red Cross partnered with Capgemini, which had been involved with the development of BabelSpeak, an AI-driven interpretation solution.

AI makes translation more accessible
BabelSpeak translates between 100 different languages in four ways: speech to speech, text to text, speech to text, and text to speech. Early versions of the digital interpreter were so promising that Capgemini decided to further develop the concept. As the solution was continually refined, Telenor’s recently established AI Factory becoming a key partner to ensure the entire solution was under Norwegian control.

Because Babelspeak and the information it generates are entirely processed and operated in Norway, it can deliver services to, among others, the police, military, healthcare services, and municipalities. These sectors require strict compliance with security, data processing, and storage regulations, requirements that are shared by many other organizations and companies.

“Capgemini is committed to ‘AI for good’ – using artificial intelligence for socially beneficial purposes,” says Thordur Arnason, Vice President at Capgemini Invent and project owner for BabelSpeak. “Being understood is a human right, and BabelSpeak helps ensure safe and effective communication for vulnerable groups in society. It saves society significant costs – and contributes to avoiding misunderstandings with potentially high risk.”

Trustworthy interpreter
Such a complex field as interpretation comes with natural challenges that affected the development of BabelSpeak. For instance, there are significant differences in data availability for various languages. Some Arabic dialects did not perform as well as initially expected, an issue that was clarified via extensive testing with a language expert.

“We hypothesized that formal and technical language – such as legal or bureaucratic – would be a major challenge,” explains Arnason. “However, BabelSpeak performed very well on vocabulary-related content. Instead, variations, nuances, and dialects in sentence structure were the most difficult to get right.”

Throughout the development, we ensured that a native speaker was always on hand to review translations for quality and precision. The project team then used the results of this oversight to further refine and improve the system.

From locked in to locked out
Since the introduction of BabelSpeak at Red Cross, the non-profit has been enabled to provide meaningful translation services to vulnerable groups. Primary among them have been refugees and formerly incarcerated individuals.

“BabelSpeak gives a voice to those sitting at the bottom of the table. Language barriers are especially challenging for refugees with experiences that make it difficult to trust authorities and interpreters. For them, BabelSpeak feels safer,” says Jeanette Steig Lid, Key Account Manager at the Norwegian Red Cross.

The Red Cross runs the “Network After Imprisonment” program, which aims to help those who want to leave a criminal lifestyle. Convicted criminals often experience the reality that there is a short path from being locked up (prison) to becoming locked out (of society) upon their release. At the same time, personal finances can be challenging.

Many also have a great need to manage their finances, build a network, and build a normal life in society. The network offers a range of activities to support this transition and effective communication is essential to success in this area.

BabelSpeak provides these groups with a new resource that offers critical services, ensuring that nobody has to face life-changing events without a crippling linguistic barrier adding to their challenges.

The collaboration with the Red Cross gave the BabelSpeak team the opportunity for extensive testing, giving them access to valuable experience and insights that would otherwise have been inaccessible. Both parties look forward to continuing the collaboration on a larger scale."
https://www.capgemini.com/news/client-stories/supporting-marginalized-communities-with-a-digital-interpreter/
#Metaglossia
#metaglossia_mundus
#métaglossie

The Department would like to remind the public to be careful when going for medical treatment and needing translation services at hospitals abroad.

 

"On January 23, 2026, the Department of Anti-Technology Crime launched an investigation and crackdown on a case of money fraud under the guise of pretending to be a translator for victims who were patients going abroad for medical treatment and deceiving victims into transferring money to their bank accounts.

 

With the coordination of the Phnom Penh Municipal Court Prosecutor, the department arrested 1 suspect, named J.L.N, a female, Cambodian.

 

Through investigation, authorities determined that the suspect used her bank account to defraud the victim.

 

The suspect has currently been referred by the Case Development Department to the Phnom Penh Municipal Court for legal proceedings.

 

The Department would like to remind the public to be careful when going for medical treatment and needing translation services at hospitals abroad.

 

Before making payment for medical treatment, please confirm and check clearly to avoid scammers who forge hospital documents to transfer money to bank accounts from unknown sources, which can eventually lead to loss of property."

https://www.khmertimeskh.com/501833209/fraudulent-medical-translator-arrested/

#Metaglossia 

#metaglossia_mundus 

#métaglossie 

"Upcoming Translation Events (Virtual & In-Person): February 2026

 

Tuesday, February 10:

 

The International Library Presents Cristina Rivera Garza on Autobiography of Cotton with Rita Indiana | Pulitzer Prize-winning author Cristina Rivera Garza discusses her new novel, Autobiography of Cotton, translation by Christina MacSweeney. In-person. Hosted by The Center for the Art of Translation...

 Starts at 7:00 p.m. (ET).

 

Thursday, February 12:

 

The 2026 Albertine Translation Prize Ceremony | On February 12, the annual Albertine Translation Prize Ceremony returns, honoring translators and American publishers of English translations of contemporary French works. This year Villa Albertine is partnering with the Colloquy series at World Poetry Books to celebrate the art of translation in a new and exciting way. Bringing together translators and readers, the evening will feature translation jousts: two translators and a moderator will engage in a lively discussion of their different renderings of the same French texts, one a work of fiction, the other a poem. In-person. Hosted by Albertine. Starts at 6:00 p.m. (ET).

 

Wednesday, February 18:

 

Colloquy #20: Translating Theory, Translators in Conversation | An event series presented by World Poetry Books in collaboration with Montez Press Radio and partnering New York City institutions and bookstores. Translators engage with live audiences in an exploration of the art of translation. For Colloquy #20, Patrick Lyons and Paul Reitter will be in conversation. In-person. Hosted by World Poetry Books. More info here. Starts at 7:00 p.m. (ET).

 

Thursday, February 26:

 

The Mushroom Gatherer: A Conversation with Author & Translator | Book launch event for The Mushroom Gatherer with readings and a moderated conversation with the author, Viktorie Hanišová, and translator, Véronique Firkusny, followed by a reception with refreshments and book signing. In-person. Hosted by Czech Center New York. More info forthcoming. Starts at 6:30 p.m. (ET)

 

Sanderling: An Evening with Anne Weber, Neil Blackadder and Tess Lewis | Deutsches Haus at NYU and Goethe Institut New York presents a reading by Anne Weber from her book Sanderling (Indigo Press, 2025, translated by Neil Blackadder), followed by a conversation between her and the book's translator Neil Blackadder, which will be moderated by the author and literary translator Tess Lewis. The conversation will focus on Anne Weber’s deeply personal reckoning with what it means to be German, how the past continues to manifest in the present, how this “Zeitreisetagebuch” (Diary of Time Travel) came to be, and finally, how the translation from German to English was facilitated by the author and translator. In-person. Hosted by Deutsches Haus at NYU. More info here. Starts at 6:00 p.m. (ET).

 

If you have an upcoming literary translation event and you'd like us to feature it on our website, please fill out this form.

https://arts.columbia.edu/content/upcoming-translation-events-virtual-person-february-2026

#Metaglossia 

#metaglossia_mundus 

#métaglossie 

"Bible society wants indigenous languages as medium of instruction in schools

Mr Sanusi said the abolition of indigenous languages as a medium of instruction in schools could undermine cultural identity.

 

The Bible Society of Nigeria (BSN) has called on the federal government to reconsider the abolition of indigenous languages as a medium of instruction in schools.

 

The general secretary/chief executive officer of BSN, Samuel Sanusi, made the call in Lagos on Saturday during a news conference to announce activities for the society’s 60th anniversary on February 8, 2026.

 

Mr Sanusi said the abolition of indigenous languages as a medium of instruction in schools could undermine cultural identity.

 

“This issue is of concern to the BSN because it has spent six decades making the word of God available and affordable to Nigerians in their preferred languages and formats,” Mr Sanusi said.

 

On the anniversary celebration, Mr Sanusi said the theme of the event would be “Celebrating Impact and Building a Legacy of Hope.”

 

According to Mr Sanusi, BSN is a member of the United Bible Societies, a global fellowship of 155 national Bible societies operating in over 200 countries and territories.

 

He disclosed that BSN was currently working on 11 Bible translations and revision projects at different stages of completion across Nigeria.

 

The BSN general secretary said the anniversary programme would begin with a Bible exhibition on February 2, 2026, at the Lagos Bible Guest House, Palmgrove, Ilupeju.

 

Mr Sanusi said, ”There will be a Bible walk on February 3 from the National Stadium, Surulere, to Obanikoro.

 

“Partners’ appreciation dinner and the dedication of a second studio for the deaf Bible translation project in Ibadan would take place on February 4.

 

“Foreign guests will arrive on February 5 for a CEOs’ conference scheduled to coincide with the anniversary celebration.”

 

He said the Founders’ Day Lecture and Awards Ceremony would be held on February 6, to be chaired by former President Goodluck Jonathan, with Pastor Poju Oyemade as the guest speaker.

 

Mr Sanusi added that a thanksgiving service would be held on February 8, 2026, at The Covenant Nation, Lagos.

 

He also announced that BSN had produced a 13-episode documentary on its activities, to be aired on DOVE TV from late January or early February.

 

The BSN boss said a commemorative book titled “Six Decades of Impact: Transmitting the WORD, Transforming Lives” would also be publicly presented.

 

He disclosed that between 2014 and 2023, BSN translated and produced 222 chronological Bible stories in Nigerian sign language for the deaf community.

 

According to him, BSN spent over N105.3 million on the Sign Language Bible Project between 2023 and 2024.

 

Mr Sanusi said BSN distributed 7,870,296 copies of assorted Bibles in the last five years.

 

“According to the United Bible Societies’ 2024 Scripture Distribution Report, BSN accounts for 1.15 million out of the 5.45 million full Bibles distributed across Africa,” the BSN boss explained.

 

Mr Sanusi disclosed that the Macedonian Call project, launched in 2018, had benefited over 50,000 people in more than 20 IDP camps nationwide.

 

He said BSN planned to spend over N306 million on the project in 2026, subject to donor support.

(NAN)"

News Agency of Nigeria • January 24, 2026

https://gazettengr.com/bible-society-wants-indigenous-languages-as-medium-of-instruction-in-schools/

#Metaglossia 

#metaglossia_mundus 

#métaglossie 

"How to Build a Neural Machine Translation System for a Low-Resource Language

An introduction to neural machine translation

 

Kaixuan Chen

Jan 24, 2026

 

In the wake of the AI boom, the pace of technological iteration has reached an unprecedented level. Previous obstacles now seem to have viable solutions. This article serves as an “NMT 101” guide. While introducing our project, it also walks readers step by step through the process of fine-tuning an existing translation model to support a low-resource language that is not included in mainstream multilingual models.

 

Background: Dongxiang as a Low-Resource Language

Dongxiang is a minority language spoken in China’s Gansu Province and is classified as vulnerable by the UNESCO Atlas of the World’s Languages in Danger. Despite being widely spoken in local communities, Dongxiang lacks the institutional and digital support enjoyed by high-resource languages. Before diving into the training pipeline, it helps to briefly understand the language itself. Dongxiang, as its name suggests, is the mother tongue of the Dongxiang people. Descended from Central Asian groups who migrated to Gansu during the Yuan dynasty, the Dongxiang community has linguistic roots closely tied to Middle Mongol. From a writing-system perspective, Dongxiang has undergone a relatively recent standardization. Since the 1990s, with governmental promotion, the language has gradually adopted an official Latin-based orthography, using the 26 letters of the English alphabet and delimiting words by whitespace.

 

Although it is still classified under the Mongolic language family, due to the prolonged coexistence with Mandarin-speaking communities through history, the language has a trove of lexical borrowing from Chinese (Mandarin). Dongxiang exhibits no overt tense inflection or grammatical gender, which may be an advantage to simplify our model training.

 

Based on the Dongxiang dictionary, approximately 33.8% of Dongxiang vocabulary items are of Chinese origin.

 

Further background on the Dongxiang language and its speakers can be found on our website, which hosts an official English-language introduction released by the Chinese government.

 

Our Model: How to Use the Translation System

We build our translation system on top of NLLB-200-distilled-600M, a multilingual neural machine translation model released by Meta as part of the No Language Left Behind (NLLB) project. We were inspired by the work of David Dale. However, ongoing updates to the Transformers library have made the original approach difficult to apply. In our own trials, rolling back to earlier versions (e.g., transformers ≤ 4.33) often triggered conflicts with other dependencies. In light of these constraints, we provide a full list of libraries in our project’s GitHub requirements.txt for your reference.

 

Our model was fine-tuned on 42,868 Dongxiang–Chinese bilingual sentence pairs. The training corpus combines publicly available materials with internally curated resources provided by local government partners, all of which were processed and cleaned in advance. Training was conducted using Adafactor, a memory-efficient optimizer well suited to large transformer models. With the distilled architecture, the full fine-tuning process can be completed in under 12 hours on a single NVIDIA A100 GPU. All training configurations, hyperparameters, and experimental settings are documented across two training Jupyter notebooks. Rather than relying on a single bidirectional model, we trained two direction-specific models to support Dongxiang–Chinese and Chinese–Dongxiang translation. Since NLLB is already pretrained on Chinese, joint training under data-imbalanced conditions tends to favor the easier or more dominant direction. As a result, performance gains on the low-resource side (Dongxiang) are often limited. However, NLLB does support bidirectional translation in a single model, and a straightforward approach is to alternate translation directions at the batch level.

 

Here are the links to our repository and website.

 

GitHub Repository

GitHub-hosted website

 

The model is also publicly available on Hugging Face.

 

Chinese → Dongxiang

Dongxiang → Chinese

 

Model Training: Step-by-Step Reproducible Pipeline

Before following this pipeline to build the model, we assume that the reader has a basic understanding of Python and fundamental concepts in natural language processing. For readers who may be less familiar with these topics, Andrew Ng’s courses are a highly recommended gateway. Personally, I also began my own journey to this field through his course.

 

Step 1: Bilingual Dataset Processing

The first stage of model training focuses on constructing a bilingual dataset. While parallel corpora for major languages can often be obtained by leveraging existing web-scraped resources, Dongxiang–Chinese data remains difficult to acquire. To support transparency and reproducibility, and with consent from the relevant data custodians, we have released both the raw corpus and a normalized version in our GitHub repository. The normalized dataset is produced through a straightforward preprocessing pipeline that removes excessive whitespace, standardizes punctuation, and ensures a clear separation between scripts. Dongxiang text is restricted to Latin characters, while Chinese text contains only Chinese characters.

Below is the code used for preprocessing:

 

import re

import pandas as pd

 

def split_lines(s: str):

    if "\\n" in s and "\n" not in s:

        lines = s.split("\\n")

    else:

        lines = s.splitlines()

    lines = [ln.strip().strip("'").strip() for ln in lines if ln.strip()]

    return lines

 

def clean_dxg(s: str) -> str:

    s = re.sub(r"[^A-Za-z\s,\.?]", " ", s)

    s = re.sub(r"\s+", " ", s).strip()

    s = re.sub(r"[,.?]+$", "", s)

    return s

 

def clean_zh(s: str) -> str:

    s = re.sub(r"[^\u4e00-\u9fff，。？]", "", s)

    s = re.sub(r"[，。？]+$", "", s)

    return s

 

def make_pairs(raw: str) -> pd.DataFrame:

    lines = split_lines(raw)

    pairs = []

    for i in range(0, len(lines) - 1, 2):

        dxg = clean_dxg(lines[i])

        zh = clean_zh(lines[i+1])

        if dxg or zh:

            pairs.append({"Dongxiang": dxg, "Chinese": zh})

    return pd.DataFrame(pairs, columns=["Dongxiang", "Chinese"])

In practice, bilingual sentence-level pairs are preferred over word-level entries, and excessively long sentences are split into shorter segments. This facilitates more reliable cross-lingual alignment and leads to more stable and efficient model training. Isolated dictionary entries should not be inserted into training inputs. Without surrounding context, the model cannot infer syntactic roles, or learn how words interact with surrounding tokens.

 

 

Bilingual dataset (by Author)

When parallel data is limited, a common alternative is to generate synthetic source sentences from monolingual target-language data and pair them with the originals to form pseudo-parallel corpora. This idea was popularized by Rico Sennrich, whose work on back-translation laid the groundwork for many NMT pipelines. LLM-generated synthetic data is another viable approach. Prior work has shown that LLM-generated synthetic data is effective in building translation systems for Purépecha, an Indigenous language spoken in Mexico.

 

Step 2: Tokenizer Preparation

Before text can be digested by a neural machine translation model, it must be converted into tokens. Tokens are discrete units, typically at the subword level, that serve as the basic input symbols for neural networks. Using entire words as atomic units is impractical, as it leads to excessively large vocabularies and rapid growth in model dimensionality. Moreover, word-level representations struggle to generalize to unseen or rare words, whereas subword tokenization enables models to compose representations for novel word forms.

 

The official NLLB documentation already provides standard examples demonstrating how tokenization is handled. Owing to NLLB’s strong multilingual capacity, most widely used writing systems can be tokenized in a reasonable and stable manner. In our case, adopting the default NLLB multilingual tokenizer (Unigram-based) was sufficient to process Dongxiang text.

 

 

Summary statistics of tokenized Dongxiang sentences (by Author)

Whether the tokenizer should be retrained is best determined by two criteria. The first is coverage: frequent occurrences of unknown tokens (<unk>) indicate insufficient vocabulary or character handling. In our sample of 300 Dongxiang sentences, the <unk> rate is zero, suggesting full coverage under the current preprocessing. The second criterion is subword fertility, defined as the average number of subword tokens generated per whitespace-delimited word. Across the 300 samples, sentences average 6.86 words and 13.48 tokens, corresponding to a fertility of approximately 1.97. This pattern remains consistent across the distribution, with no evidence of excessive fragmentation in longer sentences.

 

Overall, NLLB demonstrates robust behavior even on previously unseen languages. As a result, tokenizer retraining is generally unnecessary unless the target language employs a highly unconventional writing system or even lacks Unicode support. Retraining a SentencePiece tokenizer also has implications for the embedding layer. New tokens start without pretrained embeddings and must be initialized using random values or simple averaging.

 

Step 3: Language ID Registration

In practical machine translation systems such as Google Translate, the source and target languages must be explicitly specified. NLLB adopts the same assumption. Translation is governed by explicit language tag, referred to as src_lang and tgt_lang, determining how text is encoded and generated within the model. When a language falls outside NLLB’s predefined scope, it must first be explicitly registered, along with a corresponding expansion of the model’s embedding layer. The embedding layer maps discrete tokens into continuous vector representations, allowing the neural network to process and learn linguistic patterns in a numerical form.

 

In our implementation, a custom language tag is added to the tokenizer as an additional special token, which assigns it a unique token ID. The model’s token embedding matrix is then resized to accommodate the expanded vocabulary. The embedding vector associated with the new language tag is initialized from a zero-centered normal distribution with a small variance, scaled by 0.02. If the newly introduced language is closely related to an existing supported language, its embedding can often be trained on top of the existing representation space. However, linguistic similarity alone does not guarantee effective transfer learning. Differences in writing systems can affect tokenization. A well-known example is Moldovan, which is linguistically identical to Romanian but is written in the Latin script, while it is written in Cyrillic in the so-called Pridnestrovian Moldavian Republic. Despite the close linguistic relationship, the difference in script introduces distinct tokenization patterns.

 

The code used to register a new language is presented here.

 

def fix_tokenizer(tokenizer, new_lang: str):

    old = list(tokenizer.additional_special_tokens)

    if new_lang not in old:

        tokenizer.add_special_tokens(

            {"additional_special_tokens": old + [new_lang]})

    return tokenizer.convert_tokens_to_ids(new_lang)

 

fix_tokenizer(tokenizer,"sce_Latn")

# we register Dongxiang as sce_Latn, and it should append to the last

# output 256204

 

print(tokenizer.convert_ids_to_tokens([256100,256204]))

print(tokenizer.convert_tokens_to_ids(['lao_Laoo','sce_Latn']))

# output 

['lao_Laoo', 'sce_Latn']

[256100, 256204]

 

model = AutoModelForSeq2SeqLM.from_pretrained("facebook/nllb-200-distilled-600M")

model.resize_token_embeddings(len(tokenizer))

new_id = fix_tokenizer(tokenizer, "sce_Latn")

embed_dim = model.model.shared.weight.size(1)

model.model.shared.weight.data[new_id] = torch.randn(embed_dim) * 0.02

Step 4: Model Training

We fine-tuned the translation model using the Adafactor optimizer, a memory-efficient optimization algorithm designed for large-scale sequence-to-sequence models. The training schedule begins with 500 warmup steps, during which the learning rate is gradually increased up to 1e-4 to stabilize early optimization and avoid sudden gradient spikes. The model is then trained for a total of 8,000 optimization steps, with 64 sentence pairs per optimization step (batch). The maximum sequence length is set to 128 tokens, and gradient clipping is applied with a threshold of 1.0.

 

We initially planned to adopt early stopping. However, due to the limited size of the bilingual corpus, nearly all available bilingual data was used for training, leaving only a dozen-plus sentence pairs reserved for testing. Under these conditions, a validation set of sufficient size was not available. Therefore, although our GitHub codebase includes placeholders for early stopping, this mechanism was not actively used in practice.

 

Below is a snapshot of the key hyperparameters used in training.

 

optimizer = Adafactor(

    [p for p in model.parameters() if p.requires_grad],

    scale_parameter=False,

    relative_step=False,

    lr=1e-4,

    clip_threshold=1.0,

    weight_decay=1e-3,

)

 

batch_size = 64

max_length = 128

training_steps = 8000

warmup_steps = 500

It is also worth noting that, in the design of the loss function, we adopt a computationally efficient training strategy. The model receives tokenized source sentences as input and generates the target sequence incrementally. At each step, the predicted token is compared against the corresponding reference token in the target sentence, and the training objective is computed using token-level cross-entropy loss.

 

loss = model(**x, labels=y.input_ids).loss

# Pseudocode below illustrates the underlying mechanism of the loss function

for each batch:

 

    x = tokenize(source_sentences) # input: source language tokens

    y = tokenize(target_sentences) # target: reference translation tokens

 

    predictions = model.forward(x) # predict next-token distributions

    loss = cross_entropy(predictions, y) # compare with reference tokens

 

    backpropagate(loss)

    update_model_parameters()

This formulation actually carries an implicit assumption: that the reference translation represents the single correct answer and that the model’s output must align with it token by token. Under this assumption, any deviation from the reference is treated as an error. Even when a prediction conveys the same idea using different wording, synonyms, or an altered sentence structure.

 

The mismatch between token-level supervision and meaning-level correctness is particularly problematic in low-resource and morphologically flexible languages. At the training stage, this issue can be alleviated by relaxing strict token-level alignment and treating multiple paraphrased target sentences as equally valid references. At the inference stage, instead of selecting the highest-probability output, a set of candidate translations can be generated and re-ranked using semantically informed criteria (e.g., chrF).

 

Step 5: Model Evaluation

Once the model is built, the next step is to examine how well it translates. Translation quality is shaped not only by the model itself, but also by how the translation process is configured at inference time. Under the NLLB framework, the target language must be explicitly specified during generation. This is done through the forced_bos_token_id parameter, which anchors the output to the intended language. Output length is controlled through two parameters. The first is the minimum output allowance (a), which guarantees a baseline number of tokens that the model is allowed to generate. The second is a scaling factor (b), which determines how the maximum output length grows in proportion to the input length. The maximum number of generated tokens is set as a linear function of the input length, computed as a + b × input_length. In addition, max_input_length limits how many input tokens the model reads.

 

This function powers the Dongxiang → Chinese translation.

 

import torch

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

 

device = "cuda" if torch.cuda.is_available() else "cpu"

MODEL_DIR3 = "/content/drive/MyDrive/my_nllb_CD_model"

tokenizer3 = AutoTokenizer.from_pretrained(MODEL_DIR3)

model3 = AutoModelForSeq2SeqLM.from_pretrained(MODEL_DIR3).to(device)

model3.eval()

 

def translate3(text, src_lang="zho_Hans", tgt_lang="sce_Latn",

               a=16, b=1.5, max_input_length=1024, **kwargs):

    tokenizer3.src_lang = src_lang

    inputs = tokenizer3(text, return_tensors="pt", padding=True,

                        truncation=True, max_length=max_input_length).to(model3.device)

    result = model3.generate(

        **inputs,

        forced_bos_token_id=tokenizer3.convert_tokens_to_ids(tgt_lang),

        max_new_tokens=int(a + b * inputs.input_ids.shape[1]),

        **kwargs

    )

    outputs = tokenizer3.batch_decode(result, skip_special_tokens=True)

    return outputs

Model quality is then assessed using a combination of automatic evaluation metrics and human judgment. On the quantitative side, we report standard machine translation metrics such as BLEU and ChrF++. BLEU scores were computed using standard BLEU-4, which measures word-level n-gram overlap from unigrams to four-grams and combines them using a geometric mean with brevity penalty. ChrF++ was calculated over character-level n-grams and reported as an F-score. It should be noted that the current evaluation is preliminary. Due to limited data availability at this early stage, BLEU and ChrF++ scores were computed on only a few dozen held-out sentence pairs. Our model achieved the following results:

 

Dongxiang → Chinese (DX→ZH)

BLEU-4: 44.00

ChrF++: 34.3

 

Chinese → Dongxiang (ZH→DX)

BLEU-4: 46.23

ChrF++: 59.80

 

BLEU-4 scores above 40 are generally regarded as strong in low-resource settings, indicating that the model captures sentence structure and key lexical choices with reasonable accuracy. The lower chrF++ score in the Dongxiang → Chinese direction is expected and does not necessarily indicate poor translation quality, as Chinese permits substantial surface-level variation in word choice and sentence structure, which reduces character-level overlap with a single reference translation.

 

In parallel, bilingual evaluators fluent in both languages reported that the model performs reliably on simple sentences, such as those following basic subject–verb–object structures. Performance degrades on longer and more complex sentences. While these results are encouraging, they also indicate that further improvement is still required.

 

Step 6: Deployment

At the current stage, we deploy the project through a lightweight setup by hosting the documentation and demo interface on GitHub Pages, while releasing the trained models on Hugging Face. This approach enables public access and community engagement without incurring additional infrastructure costs. Details regarding GitHub-based deployment and Hugging Face model hosting follow the official documentation provided by GitHub Pages and the Hugging Face Hub, respectively.

 

This script uploads a locally trained Hugging Face–compatible model.

 

import os

from huggingface_hub import HfApi, HfFolder

 

# Load the Hugging Face access token 

token = os.environ.get("HF_TOKEN")

HfFolder.save_token(token)

 

# Path to the local directory containing the trained model artifacts

local_dir = "/path/to/your/local_model_directory"

 

# Target Hugging Face Hub repository ID in the format: username/repo_name

repo_id = "your_username/your_model_name"

 

# Upload the entire model directory to the Hugging Face Model Hub

api = HfApi()

api.upload_folder(

    folder_path=local_dir,

    repo_id=repo_id,

    repo_type="model",

)

Following model release, a Gradio-based interface is deployed as a Hugging Face Space and embedded into the project’s GitHub Pages site. Compared to Docker-based self-deployment, using Hugging Face Spaces with Gradio avoids the cost of maintaining dedicated cloud infrastructure.

 

Reflection

Throughout the project, data preparation, not model training, dominated the overall workload. The time spent cleaning, validating, and aligning Dongxiang–Chinese data far exceeded the time required to fine-tune the model itself. Without local government involvement and the support of native and bilingual speakers, completing this work would not have been possible. From a technical perspective, this imbalance highlights a broader issue of representation in multilingual NLP. Low-resource languages such as Dongxiang are underrepresented not due to inherent linguistic complexity, but because the data required to support them is expensive to obtain and relies heavily on human expertise.

 

At its core, this project digitizes a printed bilingual dictionary and constructs a basic translation system. For a community of fewer than one million people, these incremental steps play an outsized role in ensuring that the language is not excluded from modern language technologies. Finally, let’s take a moment to appreciate the breathtaking scenery of Dongxiang Autonomous County!

 

 

River gorge in Dongxiang Autonomous County (by Author)

Contact

This article was jointly written by Kaixuan Chen and Bo Ma, who were classmates in the Department of Statistics at the University of North Carolina — Chapel Hill. Kaixuan Chen is currently pursuing a master’s degree at Northwestern University, while Bo Ma is pursuing a master’s degree at the University of California, San Diego. Both authors are open to professional opportunities.

 

If you are interested in our work or would like to connect, feel free to reach out:

 

Project GitHub: https://github.com/dongxiangtranslationproject

Kaixuan Chen: chenkaixuann77@gmail.com

Bo Ma: brian0517m@gmail.com

 

Written By

Kaixuan Chen"

https://towardsdatascience.com/how-to-build-a-neural-machine-translation-system-for-a-low-resource-language/

#Metaglossia 

#metaglossia_mundus 

#métaglossie 

"As a rare Irish-language translator, Timothy McKeon enjoyed steady work for European Union institutions for years. But the rise of artificial intelligence tools that can translate text and, increasingly, speech nearly instantly has upended his livelihood and that of many others in his field.

He says he lost about 70% of his income when the EU translation work dried up. Now, available work consists of polishing machine-generated translations, jobs he refuses “on principle” because they help train the software taking work away from human translators. When the edited text is fed back into the translation software, “it learns from your work.”

“The more it learns, the more obsolete you become,” he said. “You’re essentially expected to dig your own professional grave.”

While workers worldwide ponder how AI might affect their livelihoods – a topic on the agenda at the World Economic Forum in Davos this week – that question is no longer hypothetical in the translation industry. Apps like Google Translate already reduced the need for human translators, and increased adoption of generative AI has only accelerated that trend.

A 2024 survey of writing professionals by the United Kingdom’s Society of Authors showed that more than a third of translators had lost work due to generative AI, which can create sophisticated text, as well as images and audio, from users’ prompts. And 43% of translators said their income had dropped because of the technology.

In the United States, data from 2010-23 analyzed by Carl Frey and Pedro Llanos-Paredes at Oxford University showed that regions where Google Translate was in greater use saw slower growth in the number of translator jobs. Originally powered by statistical translation, Google Translate shifted to a technique called neural translation in 2016, resulting in more natural-sounding text and bringing it closer to today’s AI tools.

“Our best baseline estimate is that roughly 28,000 more jobs for translators would’ve been added in the absence of machine translation,” Frey told CNN.

“It’s not a story of mass displacement but I think that’s very likely to follow.”

Timothy McKeon refuses to edit machine translations, saying it's like digging "your own professional grave." Courtesy Timothy McKeon
The story is similar globally, suggests McKeon: He is part of the Guerrilla Media Collective, an international group of translators and communications professionals, and says everyone in the collective supplements their income with other work due to the impact of AI.

‘The entire US is looking at Wisconsin’
Christina Green is president of Green Linguistics, a provider of language services, and a court interpreter in Wisconsin.

She worries her court role could soon vanish because of a bill that would allow courts to use AI or other machine translation in civil or criminal proceedings, and in certain other cases.

Green and other language professionals have been fighting the proposal since it was introduced in May. “The entire US is looking at Wisconsin” as a precedent, Green said, noting that the bill’s opponents had so far succeeded in stalling it.

While Green still has her court job, her company recently lost a major Fortune 10 corporate client, which she said opted to use a company offering AI translation instead. The client accounted for such an outsized share of her company’s business that she had to make layoffs.

Christina Green has had to let staff go because her translation company has lost a large amount of work to AI. Courtesy Alvin Connor/Havone Studios
“People and companies think they’re saving money with AI, but they have absolutely no clue what it is, how privacy is affected and what the ramifications are,” Green said.

‘Governments are not doing enough’
Fardous Bahbouh, based in London, is an Arabic-language translator and interpreter for international media organizations, including CNN. She has seen a considerable reduction in written work in recent years, which she attributes to technological developments and the financial pressures facing media outlets.

Bahbouh is also studying for a PhD focusing on the translation industry. Her research shows that technology, including AI, is “hugely impacting” translators and interpreters.

Governments should be doing more to protect foreign-language professionals from the threat posed by AI, according to Fardous Bahbouh. Courtesy Fardous Bahbouh
“I worry a great deal that governments are not doing enough to help them transition into other work, which could lead to greater inequality, in-work poverty and child poverty,” she told CNN.

Many translators are indeed looking to retrain “because translation isn’t generating the income it previously did,” according to Ian Giles, a translator and chair of the Translators Association at the UK’s Society of Authors. The picture is similar in the United States: Many translators are leaving the profession, Andy Benzo, president of the American Translators Association, told CNN.

And Kristalina Georgieva, the head of the International Monetary Fund, said in Davos Thursday that the number of translators and interpreters at the fund had gone down to 50 from 200 due to greater use of technology.

Governments should also do more for those remaining in the translation industry, by introducing stronger labor protections, Bahbouh argued.

Human professionals still needed
Despite advances in machine translation and interpretation, technology can’t replace human language workers entirely just yet.

Andy Benzo, president of the American Translators Association, says the risks of using AI translation in "high-stakes" fields are "humongous." Courtesy Andy Benzo
While using AI tools for everyday tasks like finding directions is “low-risk,” human translators will likely need to be involved for the foreseeable future in diplomatic, legal, financial and medical contexts where the risks are “humungous,” according to Benzo.

“I’m a translator and a lawyer and in both professions the nuance of each word is very specific and the (large language models powering AI tools) aren’t there yet, by far,” she said...

Giles, who translates commercial fiction from Scandinavian languages into English, used to supplement his income with translation work from companies, but that has now disappeared. Meanwhile, literary commissions have continued to come in, he said.

There’s also one key element of communication that AI can’t replace, according to Oxford University’s Frey: Human connection.

“The fact that machine translation is pervasive doesn’t mean you can build a relationship with somebody in France without speaking a word of French,” he said."
Lianne Kolirin
https://edition.cnn.com/2026/01/23/tech/translation-language-jobs-ai-automation-intl
#Metaglossia
#metaglossia_mundus
#métaglossie

"Iowa State University researchers received unexpected answers to the question of how often people humanize artificial intelligence in news writing.

 

ISU English professor Jo Mackiewicz and Jeanine Aune, an English teaching professor and director of the university’s advanced communication program, recently published a study on how prevalent the use of anthropomorphizing, or humanizing, language is when it comes to artificial intelligence programs.

 

They were surprised to find that AI was not typically described in human terms. Even the use of humanizing verbs like “learns,” when considered in context, did not always treat the technology like a person.

 

“It’s the exact opposite of what I was expecting,” Aune said.

 

Aune said the pair got the idea for this research from a conference they both attended. The discussion included the suggestion that educators emphasize that AI is a tool that cannot replace communication principles and practices and to avoid anthropomorphizing the technology when they talk about it.

 

Iowa State University alumni were also involved in the study, including current Brigham Young University associate professor of linguistics Matthew Baker and University of Northern Colorado assistant professor of English Jordan Smith.

 

Heading into their research, Aune and Mackiewicz said they both had their assumptions as to what they’d find — that humanizing the technology would be widespread. Prior research has shown people anthropomorphize when working with robotics, and Aune said they’ve read articles written about how people use and perceive AI that have suggested humanization is happening.

 

The research team used News on the Web to study language relating to AI, Mackiewicz said, as it includes the most recent data aggregated from news articles originating from 20 different countries. The database has topped 20 billion words, and it includes some of the earliest news articles about AI to recent articles covering the technology.

 

“When we started to do an analysis of the individual pairings, that’s where the findings became most interesting, because the overall finding was that, oh, they don’t really pair as frequently as prior research or opinion pieces would have you think,” Mackiewicz said.

 

Words the team was on the lookout for focused on “verbs that reflect cognition,” Mackiewicz said, also known as “mental verbs” — needs, learns, means, and understands are just a few examples. They also narrowed the search to references to AI or ChatGPT, as it was one of the first AI tools to become publicly available and known.

 

What they found is that these mental verbs are not often paired with the identified terms, and there was nuance to the instances where they were used together.

 

“AI means, or AI needs, or ChatGPT knows — you put those together, just on the surface, they seem anthropomorphizing, but they’re not necessarily,” Mackiewicz said.

 

Anthropomorphization “exists on a spectrum,” Mackiewicz said, where a word like “means” could apply to different senses and meanings depending on how it is being used. Other words have different meanings in their common usage or usage by different disciplines, such as “learns.” Mackiewicz said the terms “AI” and forms of “learn” were paired together frequently in the data, which to some would seem to say that the technology is learning like a human could, but to others would mean something completely different.

 

It’s this nuance that shows why strict guidelines of what words can and cannot be used in relation to AI are flawed, she said, and what advice the researchers would offer instead is that people need to be careful about knowing who their audience is when writing about AI and thinking about how their words could be interpreted.

 

Mackiewicz said that in a way, this work showed the value of human beings. It’s easy for a computer to count how many times words are paired together, but it’s harder to analyze it within its context and determine the language’s actual meaning.

 

While it would be a bigger lift than this study, the researchers said they would be interested in studying what language is used to refer to AI in other areas of writing, like social media. This would require them to create their own database and scrape social media for the necessary information, Mackiewicz said, but it could yield interesting results.

 

While journalists utilize style guides created by the Associated Press or other organizations when deciding how to write on certain topics and receive editing on their writing, Mackiewicz said, social media generally doesn’t follow such rules.

 

“We now even had some students refer to ChatGPT with a male pronoun, not talk about it as ‘it,’” Aune said. “So I personally really advocate being intentional with the language, so talking about output text and it uptakes the prompt, you could really kind of emphasize that it is a tool that’s really helpful, but it’s not replacing our brains.”"

https://iowacapitaldispatch.com/2026/01/23/iowa-state-university-researchers-dive-into-language-that-humanizes-ai-systems/

#Metaglossia 

#metaglossia_mundus 

#métaglossie 

"The Language Flagship Program at Brigham Young University was discontinued... because of government funding cuts.

The program, funded by the Department of Defense, was an undergraduate program that offered students opportunities to enhance their language skills to a professional level.

According to the flagship center website, students were provided with experiences that focused on global engagement, professional skills, language proficiency and community and service. At BYU, Chinese and Arabic were offered through the program.

Kirk Belnap, a professor in the Department of Asian and Near Eastern Languages, was the director of the flagship program.

Belnap said a loss associated with the cancellation of the program was the departure of Ahmad Karout, a native Arabic speaker from Syria who worked as the academic director and coordinator for the program.

“We don’t have an Arab who is full-time on the faculty, so losing Ahmad Karout was a real loss for us,” Belnap said.

One of the opportunities offered through the program was the chance to travel abroad for intensive language study.

Chinese Flagship students were sent to China before COVID-19 and then to Taiwan post-pandemic. The Arabic study abroad was in Morocco.

Nicholas Heil joined the Arabic flagship program in 2020 and participated in the capstone from 2023 to 2024 before its discontinuation.

“It provided a lot of opportunities that were really cool, and then the capstone was really helpful for getting to an advanced level of Arabic,” Heil said.

“I think we were cut because we are a private university and they knew that they could cut us and we’d be okay," Belnap said.

Even without this program available, BYU has created an independently funded Master of Arts in Professional Language (MAPL). This program provides similar opportunities for students.

The MAPL program first admitted students learning Chinese in fall 2022. It began teaching Arabic in 2024 and administrators hope to offer French sometime this year. Steve Riep is the director of the MAPL program and has been with the program since it started.

While the flagship program was more flexible, Riep said the MAPL program requires a professional focus.

“We give them an opportunity to develop professional-level fluency in a target language, in a professional domain,” Riep said. “So it would be in whatever their professional area is. That could be microbiology, engineering, business or even diplomacy.”

A study abroad is also offered through MAPL, allowing students to travel to Taiwan and Morocco to take a specified profession class in their intended language. Students also complete an internship abroad as a part of their education.

“The program is growing; it’s begun sort of small, but our hope is that this year we’re going to have a much larger pool of applicants than we’ve had in the past. We’re excited about the future,” Riep said."
Sophia Howcroft,
January 23, 2026
https://universe.byu.edu/campus/byu-cuts-arabic-chinese-flagship-programs-following-federal-funding-loss
#Metaglossia
#metaglossia_mundus
#métaglossie

"Bill would protect language access for 25 million individuals in the U.S. with limited English proficiency, including 32 percent of Asian Americans and 12 percent of Native Hawaiians and Pacific Islanders.

WASHINGTON, D.C. — Today, Chair of the Congressional Asian Pacific American Caucus (CAPAC) Rep. Grace Meng (NY-06), Chair Emerita Rep. Judy Chu (CA-28), and CAPAC members Rep. Dan Goldman (NY-10) and Rep. Juan Vargas (CA-51) introduced the Language Access for All Act of 2026 to codify language access requirements for federal agencies, including translation and interpretation services under threat from the Trump administration.

 

In March 2025, President Trump signed Executive Order (EO) 14224 that declared English as the official language of the United States and revoked EO 13166, a 25-year-old mandate that required agencies and recipients of federal funding to provide meaningful language access to individuals with limited English proficiency (LEP). The Trump administration's Department of Justice issued new guidance that minimizes multilingual services and redirects resources towards English language education and assimilation.

 

These policy changes threaten language access for the over 25 million individuals in the United States—eight percent of the U.S. population—with limited English proficiency. Asian Americans have among the highest language access needs of any racial group, with 32 percent having LEP. 12 percent of Native Hawaiians and Pacific Islanders also have significant language access needs. And while Spanish language speakers make up the majority of those who speak another language in the United States, nearly 40 percent report speaking English “less than very well” in the most recent U.S. Census.

 

The Language Access for All Act of 2026 modernizes and strengthens the federal government’s approach to language access by codifying EO 13166 and establishing a coordinated, accountable framework to ensure meaningful access for individuals with limited English proficiency. The legislation promotes consistency across agencies, increases transparency and public engagement, and updates federal language access policy to reflect evolving technologies.

 

“Every American deserves equal access to federal services and programs in a language they can understand. Language access is essential to ensure individuals are able to access small business loans or receive the right medical care,” said Rep. Grace Meng, Chair of the Congressional Asian Pacific American Caucus. “I am proud to introduce the Language Access for All Act alongside my colleagues to safeguard translation services for individuals with limited English proficiency, including millions in the Asian American, Native Hawaiian, and Pacific Islander community. We will continue to fight against the Trump administration's attacks on immigrants and the essential services that our communities rely on and deserve.”

 

“For more than 25 years, both Democratic and Republican presidents have supported language accessibility across the federal government. Trump’s roll back of these protections is simply wrong. In my district, translation services are essential for parents applying for a home loan, seniors accessing Medicare, immigrants starting a small business, and disaster survivors accessing the FEMA’s resources. That is why I’m proud to co-lead the Language Access for All Act to ensure no one is denied health care, housing, or disaster assistance because English is not their first language. Language access is a civil right and rolling back these services is an attack on our immigrant communities,” said Congressmember Judy Chu (CA-28), CAPAC Chair Emerita.

 

“In my district, in the most linguistically diverse city on earth, language access mandates across the federal government helps ensure that everyone has meaningful access to housing loans, health care, workforce programs, and life-saving emergency alerts. Trump's attempt to roll back language access is just one of his many xenophobic attempts to attack immigrant communities, and it is completely unacceptable. I am proud to be introducing legislation to restore common sense requirements and ensure information about basic government services is made available to all,” said Rep. Dan Goldman.

 

“For decades, federal language access services have helped millions of people file taxes, get emergency alerts, apply for loans, and access health care. Trump’s decision to designate English as our country’s official language and attempt to scrap these critical services is dead wrong,” said Rep. Juan Vargas. “No one should be locked out of federal programs because of the language they speak. This legislation is critical as we fight to push back on Trump’s anti-immigrant agenda and keep in place the services our communities rely on.”

 

The Language Access for All Act of 2026:

 

Requires federal agencies to ensure that individuals with LEP can meaningfully access the federally conducted programs and activities of the agency, including through translation and interpretation.

Creates a public complaint system to track complaints regarding barriers to meaningful access at agencies.

Requires agencies to develop and maintain language access plans consistent with EO 13166, with public notice and comment, and to submit plans to Congress and publish them on LEP.gov.

Establishes language access technical standards that allow individuals with LEP to access agency content and applies to all agency communications, including AI and automated language assistance services.

Ensures AI-assisted language services do not replace qualified translators and interpreters, comply with federal privacy requirements, and are continuously tested for bias, discrimination, and errors.

Creates an interagency language access working group to provide guidance, coordination, and technical assistance.

Requires each agency to designate a language access coordinator to lead implementation and serve as a point of contact.

Together, these reforms aim to improve service delivery, reduce barriers to access, and ensure federal agencies are equipped to meet the language needs of the public.

 

This legislation builds on previous actions from CAPAC lawmakers to protect language access.

 

In April 2025, Members demanded answers from the 15 federal agencies regarding potential disruptions to services and programs for individuals with limited English proficiency, and demanded answers from President Trump and the Department of Justice about the steps the administration is taking to ensure essential translation services continue uninterrupted.

 

In July 2025, CAPAC leadership issued a joint statement condemning the Department of Justice’s new memorandum that slashes access to multilingual government services.

 

The Language Access for All Act of 2026 is endorsed by 53 organizations..."

January 23, 2026

Press Release

https://capac.house.gov/press-release/meng-chu-goldman-and-vargas-introduce-bill-protect-multilingual-services-federal

#Metaglossia 

#metaglossia_mundus 

#métaglossie 

"Des chercheurs européens ont mis au point un modèle baptisé EuroLLM-22B. Complètement open source, il intègre les 24 langues officielles de l'UE et 11 autres langues.

 

Face à la dominance des LLM anglo-saxons ou chinois, le projet EuroLLM entend se démarquer à travers le nombre de langues adressées et son caractère totalement open source. Le modèle a été développé par l’Universidade de Lisboa (Instituto Superior Técnico), l’Université d’Édimbourg, l’Université Paris‑Saclay, Sorbonne Université, Naver Labs, Unbabel et l’Université d’Amsterdam. EuroLLM a également reçu le soutien des programmes européens comme Horizon Europe pour financer la recherche et l’innovation et EuroHPC dédiée au calcul haute performance. Au sein du laboratoire MICS (Mathématique, informatique et système complexe) de CentraleSupélec, deux doctorants en informatique Hippolyte Gisserot-Boukhlef et Nicolas Boizard ont développé le modèle EuroLLM-22B (22,6 milliards de paramètres), entraîné sur le supercalculateur MareNostrum 5 du Barcelona Supercomputing Center.

 

Dès sa conception, EuroLLM-22B a voulu intégrer les 24 langues officielles de l’Union européenne. Il a ajouté 11 langues supplémentaires parmi lesquelles l’arabe, le catalan, le galicien, le norvégien, le russe, le turc, l’ukrainien, le chinois, l’hindi, le japonais et le coréen. Au démarrage, le groupe universitaire EuroLLM a livré un premier modèle modeste avec 1,7 milliard, puis 9 milliards de paramètres. « Nous avons progressivement augmenté l’échelle pour relever des tâches de complexité croissante : mathématiques, code, traduction », explique Hippolyte Gisserot-Boukhlef. Aujourd’hui, le modèle atteint 22,6 milliards de paramètres. Les premiers résultats donnent de bonnes performances dans la compréhension, la traduction et la génération de texte multilingue.

 

Un ADN complétement open source

L’autre aspect sur lequel insiste le co-fondateur d’EuroLLM-22B, c’est le caractère complètement open source du LLM. « En partant d’un modèle existant, même open weight, une partie de la recette d’entrainement reste inconnue. Or pour revendiquer le full open source, il faut que tout soit transparent : poids, données et méthodologie », souligne-t-il. Une référence à la distinction entre les fournisseurs qui proposent des LLM dits ouverts comme Llama, Gemma, Qwen ou Deepseek, mais qui ne publient pas leurs données d’entraînement. EuroLLM-22B, qui a utilisé un dataset de 4 000 milliards de tokens pour son training en puissant dans les ressources de Wikipédia et Arvix (travaux de recherche universitaires), se compare à des modèles comme Olmo développé par AI2 et Apertus élaboré par un consortium d’universités suisses.

 

Pour les développements futurs d’EuroLLM, plusieurs pistes sont explorées. « Nous travaillons sur des architectures de type Mixture of Experts, qui offre la possibilité de réduire les coûts de calcul tout en maintenant un haut niveau de performance », précise Nicolas Boizard. Par ailleurs, la multimodalité (audio et vidéo) est un axe de réflexion en s’appuyant sur les capacités de calcul du supercalculateur Jupiter à partir de 2026. « L’objectif est de créer des modèles performants sur plusieurs types de données, de les combiner et d’améliorer globalement leurs résultats », conclut le co-fondateur."

Louise Costa et Pierre Khan

22 janvier 2026

https://www.lemondeinformatique.fr/actualites/lire-eurollm-22b-parie-sur-un-modele-open-source-et-multilingue-98963.html

#Metaglossia 

#metaglossia_mundus 

#métaglossie 

"Une nouvelle traduction pour "La Mort à Venise" de Thomas Mann

Plus d’un siècle après la première publication de "La Mort à Venise", une nouvelle traduction française restitue la force et la complexité d’un texte majeur, dont le trouble, dans une Venise gagnée par le choléra, demeure intact.

Avec
Claire de Oliveira, traductrice, maîtresse de conférence à l'Université Paris-Sorbonne.
Thomas Mann est né en juin 1875, il y a un peu plus de 150 ans, ce chiffre rond a occasionné plusieurs manifestations, lectures et relectures en 2025 partout dans le monde. En 2026, son œuvre entre dans le domaine public dans de nombreux pays et vient alors relancer les lectures et les adaptations de ses textes. Dans ce sens, une nouvelle traduction de « La Mort à Venise », texte publiée en 1912, paraît en 2026 avec d'autres nouvelles aux éditions Christian Bourgois. Nous y lisons la ville italienne, plusieurs mystérieuses apparitions masculines, le désir homosexuel, le vieillissement, la création et la mort. Venise y est, tour à tour somptueuse, malsaine et soumise à une épidémie de choléra.

La langue de Thomas Mann
Claire de Oliveira trouve la langue de Thomas Mann "somptueuse, très originale et en lutte contre une peur de l'inachèvement et du silence. Pour Thomas Mann , l'un des moyens d'échapper au mutisme est ce qu'il appelle l'imposture. L'incipit de « La mort à Venise » est une parodie du style d'Achenbach qui est lui-même le personnage principal de cette nouvelle."

Le grand sujet de « La mort à Venise »  pour Claire de Oliveira est "la passion interdite de l'homosexualité. À mot couvert, Thomas Mann nous parle dans ce texte de l'homosexualité pénaliséeet réprimée. L'engouement pour cet adolescent du même sexe est exprimé par des métaphores qui le subliment, qui se situent au-delà du délabrement pour s'élever vers la contemplation platonicienne de la beauté" dans laquelle Thomas Mann "élabore une conception inédite de l'esthétique."

Prendre soin du lecteur dans la traduction
Claire de Oliveira prend soin du lectorat contemporain de Thomas Mann grâce à un appareil critique comprenant "des notes explicatives" pour que "certaines allusions culturelles" puissent "être décelées par les lecteurs contemporains francophones." Claire de Oliveira a également été, dans sa traduction, attentive aux "'intertextes au sens large, car Thomas Mann est un très grand lecteur et ne cesse de faire des allusions à des auteurs qu'il adore." Il y a notamment des "références extrêmement cryptées à Gustave Flaubert."

La question de l'auditrice Yulia @yulia_morel à l'attention de Claire de Oliveira : "En tant que traductrice, comment garder quelque chose de l'esprit ou de l'identité allemande propre à Thomas Mann en français ? Faut-il forcément chercher à le préserver ou accepter que la traduction transforme notre lecture ?""
Jeudi 22 janvier 2026
Provenant du podcast
Le Book Club
https://www.radiofrance.fr/franceculture/podcasts/le-book-club/une-nouvelle-traduction-pour-la-mort-a-venise-de-thomas-mann-2557648
#Metaglossia
#metaglossia_mundus
#métaglossie