Juan Antonio Pérez

Some NLP tasks may seem solved, but modern AI excludes thousands of languages which lack massive training data 🌐 Yet for decades, field linguists or speakers have documented them in detail. How can AI leverage this wealth to bring language technologies to these languages? (1/3)

PhD opportunity! 🚨🚨 We’re hiring a 4-year predoctoral researcher to investigate unseen language acquisition in LLMs using resources such as grammar books. Join our supportive, respectful team for this challenging project.

🔗 Details and contact info: transducens.dlsi.ua.es/opening-for-a-…

. @nuriaoliver motivando a los jóvenes a saber más de la IA en un salón de actos abarrotado en la @UA_Universidad. Hablando de su trabajo en @ELLIS_Alicante y de paso haciendo un gran favor a la ciencia y también a la @UA_Universidad para tener a estudiantes motivados pronto.

Qué divertido fue participar ayer como juez de la categoría Explore en la celebración de la First LEGO League #FLL en la @UA_Universitat 🐳🫧 Gracias a los organizadores, especialmente a @d_tomas y @bigpacopujol, por invitarme 🐚🐬🪸🐙

The dataset PILAR github.com/transducens/PI… that our @transducens research group at @UA_Universidad @UA_Universitat collected and released in 2024 was included in the training data of the ALIA machine translation models langtech-bsc.gitbook.io/alia-kit/datas… for Aragonese, Aranese, and Asturian, developed by the @BSC_CNS and presented last week by the Spanish Government: alia.gob.es 🥳

According to their model cards, the translation models were developed by the @BSC_CNS as part of the Shared Task on Translation into Low-Resource Languages of Spain that we successfully organized last year: www2.statmt.org/wmt24/romance-…

Note that we have also contributed our own multilingual model, called IBRO, designed for translation from Spanish into several Romance languages, including Aragonese, Aranese, Asturian, Catalan (Valencian included), and Galician, as well as from Catalan into Aranese. The 1.3B IBRO model huggingface.co/Transducens/Ib… was publicly funded through the LiLowLa project: transducens.dlsi.ua.es/lilowla/

Although the resources invested are significant, greater support from public administrations is essential. ✊🔥 Some impactful measures, like releasing certain public linguistic resources under open licenses, are not especially costly but could have a major impact.