Tanja Samardžić , ReLDI member, University of Geneva, Programme chair

Mirjana Starović, ReLDI member, Lexicom Belgrade, Local organiser

Local advisory members:

Vlado Delić, University of Novi Sad

Sabina Halupka Rešetar, University of Novi Sad

Boško Nikolić, University of Belgrade

Ranka Stanković, University of Belgrade

Scientific committee:

Boban Arsenijević, University of Graz

Duygu Ataman, Middle East Technical University, Ankara

Vuk Batanović, University of Belgrade, ReLDI member

Christian Bentz, University of Passau

Silvia Bernardini, University of Bologna

Ksenija Bogetić Pejović, Slovenian Academy of Sciences and Arts

Milica Denić, University of Geneva

Ljiljana Dolamić, Cyber-Defence Campus, armasuisse

Tomaž Erjavec, Jožef Stefan Institute

Adriano Ferraresi, University of Bologna

Darja Fišer, University of Ljubljana and CLARIN ERIC

Goran Glavaš,  University of Würzburg

Gordana Hržica, University of Zagreb

Tihana Kraš, University of Rijeka

Nikola Ljubešić, Jožef Stefan Institute, ReLDI member

Aleksandra Miletić, University Paris Nanterre

Filip Miletić, University of Stuttgart

Maja Miličević Petrović, University of Bologna, ReLDI member

Mirjana Mirić, Serbian Academy Of Sciences and Arts

Sandra Mitrović, IDSIA Lugano

Giuseppe Samo, University of Geneva and Idiap

Yves Scherrer, University of Oslo

Milan Sečujski, University of Novi Sad

Nikola Simić, University of Novi Sad

Marko Simonović, University of Graz

Irena Srdanović, University of Pula

Miloš Stanojević, University College London

Siniša Suzić, University of Novi Sad

Kristina Štrkalj Despot, Institute of Croatian Language and Linguistics

Slavko Žitnik, University of Ljubljana

The international conference Language Science and Technology was held on 25–26 September 2025 at the Palace of Science in Belgrade, organised by the ReLDI Centre for Language Data. The conference provided an overview of current research and emerging trends in computational modelling of natural language, including generative models such as GPT and other approaches based on machine learning and large-scale linguistic data.

Particular attention was given to machine translation and speech-to-text and text-to-speech technologies in the context of less-resourced languages such as Serbian, Croatian and Slovenian, as well as their regional varieties. Across the sessions, a common theme was the need for further development of language infrastructure and for turning language technologies into practical applications.

Roundtable: How are language technologies changing language services?

The first session focused on how language technologies, particularly generative models, are transforming the language services industry, including translation and software and product localisation.

The main conclusion was that language services are becoming increasingly automated, reducing the need for specialised expertise to areas such as court interpreting and literary translation. At the same time, the growing availability of technological solutions is expanding the use of automated services such as speech transcription and synthesis. While these technologies were once limited to specific applications, such as making media content accessible to people with visual or hearing impairments, they are now being used across a wide range of areas — from business process automation to controlling devices through natural language and adapting media content for wider audiences.

This development also creates new opportunities for language professionals, who can play an important role in creating data for training and evaluating generative models. The SPARK model developed by iFLYTEK is one example: its performance is continuously improving, largely thanks to high-quality training and evaluation data. The model already offers some support for Serbian, but improving its performance requires experts in the Serbian language who are also familiar with modern computing technologies.

Main scientific programme

Sessions in the main scientific programme presented the current performance of a range of modern models on tasks involving thematic and regional text classification, as well as speech-to-text conversion.

Beyond performance measured through established scientific methods, the research presented also showed that regional characteristics of language are encoded in trained models, and that the composition of training data is particularly important for smaller models with fewer parameters. Such models can also be more effectively specialised for specific language varieties and domains.

The sessions also presented new insights into the structure and use of language based on the analysis of large-scale linguistic data. For example, one study showed that higher voice pitch is associated with negative sentiment in parliamentary speeches, while no clear patterns were observed in variations in vocal intensity. Quantitative analysis of verb paradigms showed that Torlak speech is structurally closer to Serbian than to Bulgarian, challenging previous assumptions.

Synergies

Participants in this session presented activities carried out through several projects in Serbia and the region, including the TESLA and AI-SPEAK projects funded by the Science Fund of the Republic of Serbia. These projects focus on developing and training modern generative models specifically for Serbian.

The session also presented results from the COMtext.SR project, funded by the local IT community, which focuses on developing linguistic resources and technologies for processing legal and administrative texts in Serbian.

The discussion of these projects highlighted the need for further activities to put the technologies developed through them to practical use and turn them into solutions for end users.

Roundtable: National strategies for the development of language technologies

Although modern generative models are largely multilingual and can respond to users in a wide range of languages, their performance drops significantly across all languages compared with English.

Achieving satisfactory performance for any other language requires a national strategy and government support for the development of research infrastructure for language technologies. A national strategy should coordinate and align the work of all stakeholders involved in the development of language technologies within a country and direct their efforts towards a common goal.

In addition to researchers at universities and research institutes, important stakeholders include public institutions with large archives, such as the National Library and Radio Television of Serbia (RTS), as well as end users of language technologies, including public administration, companies and individuals.

In Norway, for example, the national library plays a leading role in the development of language technology infrastructure. In the United Kingdom, cooperation on shared infrastructure is encouraged through the national science fund, which sets standards for the openness and accessibility of results produced by publicly funded projects.

While competition between researchers is encouraged in research itself, cooperation is encouraged when it comes to shared infrastructure. For Serbia, the most effective approach would be to establish a consortium of research and other institutions, including public institutions, that would jointly develop and maintain the infrastructure, with a clear division of tasks, responsibilities and obligations, while competition would continue to be encouraged in other areas of research.

A call for infrastructure investment

Just as roads and railways provide infrastructure that supports economic development as a whole, technological infrastructure for language processing can support the development and application of artificial intelligence as an emerging area of the IT industry.

For widely spoken languages, the development of large language models and their diverse applications has been driven largely by market demand and significant commercial opportunities. This is naturally not the case for less-resourced languages such as the South Slavic languages.

In this region, the development of linguistic resources and tools has so far been driven primarily by academic research. To make the transition towards commercial applications and wider use of language technologies, systemic support from society and the state is needed through investment in infrastructure.

Advances in natural language processing can improve the efficiency of public administration and government institutions — including the judiciary, ministries and local authorities — as well as public media; support the preservation of cultural heritage in libraries, museums and archives; and improve access to a range of services for citizens.

It is therefore essential for government institutions — including ministries and offices responsible for the ICT sector, as well as relevant departments of major public and private companies — to increase their contribution to and involvement in the development of language infrastructure and its practical applications.

The conference also presented a proposed strategy for the development of language technologies in Serbia. The document will be presented to the wider community in the coming period and may be further revised based on the feedback received.

The strategy is primarily intended for relevant government institutions, providing them with a concrete framework for the development of language technologies that can support decision-making and planning across all areas related to language infrastructure.Report by session: