Till konferenssidan | Conference page
ASTIN Conference
|
16:30 – 17:30 |
Poster session: Infrastructures (Foyer 1) |
|
17:30 – 19:00 |
University of Helsinki Reception (Room Christina) |
Till konferenssidan | Conference page
|
08:30 – 09:00 |
Registration |
|
09:00 – 12:00 |
Session 1: Small languages and large models |
|
Keynote: Architecting Sovereign Norwegian AI Javier de la Rosa, The National Library of Norway |
|
|
Coffee Break |
|
|
Qualitative evaluation of ASR frameworks for North Sámi using knowledge-based tools Katri Hiovain-Asikainen, UiT The Arctic University of Norway Abstract: In the presentation, three recent North Sámi automatic speech recognition (ASR) systems are compared, examining their training data, architecture, and linguistic coverage. Beyond standard error-rate metrics, rule-based text proofing tools to uncover systematic linguistic errors typical for each ASR framework can be utilized in the development process of ASR systems. Preliminary results show that large multilingual models achieve competitive error rates but often produce linguistically implausible word forms. The resulting hybrid evaluation framework can be extended to dialect-specific data for developing usable ASR for practical uses like subtitling, archiving, education, and public-sector accessibility. Additionally, future plans for Sámi speech technology are discussed in the presentation. Bio: Katri Hiovain-Asikainen works as a speech technologist in the Divvun group at the Arctic University of Norway, where she designs and leads speech technology projects for the Sámi languages. Her work includes building speech corpora and developing speech synthesis and automatic speech recognition for several Sámi languages. Divvun has published speech synthesis for North, Lule and South Sámi, and is currently working on ASR for these languages. She is a part of the AiLab/Språkmodeller team at the Norwegian National Library, developing speech technology for languages spoken in Norway. Trained in linguistics and phonetics, she received her PhD from the University of Helsinki in 2023. Her dissertation examined how Finnish and Norwegian have influenced prosody, such as quantity and intonation, in regional varieties of spoken North Sámi. She is also a visiting researcher in the Phonetics and Speech Synthesis Research Group at the University of Helsinki, where she has worked on automatic dialect recognition and on evaluating how well speech synthesis handles complex prosodic features. Her work aims to support the revitalisation of minority languages by making speech technology available to their communities. |
|
|
ASR for Northern Sámi: building speech foundation models for an underrepresented language Yaroslav Getman, Aalto University Abstract: This talk presents the challenges of building automatic speech recognition (ASR) for Northern Sámi, a language with a small speaker community across Finland, Norway and Sweden, and the solutions we have developed at Aalto University from 2024 to 2026. We first compared adaptation strategies for existing speech models when both transcribed and untranscribed Sámi data are limited. We then pretrained self-supervised Sámi speech foundation models from scratch on untranscribed radio broadcasts from KAVI’s archive. Recent experiments with encoder architectures and semi-supervised learning yielded further improvements. Overall, these steps reduced the word error rate on a diverse Sámi test set from 48% to 20%, with the largest improvements on spontaneous speech. Our models are available through two demos, including a transcription service hosted at CSC in Finland. Bio: Yaroslav Getman is a doctoral researcher in the Automatic Speech Recognition group at Aalto University. His research focuses on self-supervised speech representation learning for underrepresented languages, which lag behind English and other high-resource languages in speech technology. His work spans adapting self-supervised models in low-resource settings, developing foundation models for underrepresented languages, and interpreting what these models learn internally. |
|
|
Using LLMs to improve rules-based technologies, and other side quests Brendan Molloy, Necessary Innovations AB Abstract: Rules-based technologies are still relevant in this apparently LLM-crazy world. With Divvun, we have used Claude Code and Codex to great effect to port several dated codebases into maintainable and modern Rust, enabling more use cases than were previously possible, while also utilising LLMs to improve the effectiveness of spell checking and grammar checking. This talk will go through the process of using these tools so that you too may utilise them for your own purposes. Bio: Brendan Molloy is a software engineer and minority language supporter who has worked predominantly on Sámi language technologies and tools such as spell checking, grammar checking, keyboards and other experiments for the last decade and a half. He is both the CEO of Necessary Innovations AB and the Chair of the Institute for Open Language, and lives in Göteborg, Sweden. |
|
|
12:00 – 13:00 |
Lunch |
|
13:00 – 16:00 |
Session 2: Data, data, data |
|
Keynote: The Data We Have and the Data We Can Use: Cultural Heritage in the Age of AI Christopher Natzén, The National Library of Sweden Abstract: Data is a fundamental part of the infrastructure for AI development, particularly for smaller languages, making national libraries and cultural heritage institutions strategically important. Drawing on the experience of the National Library of Sweden, this keynote explores how legal deposit, digitisation and data stewardship can transform cultural heritage collections into usable data for research and language model development. It highlights the challenges posed by limited digitisation, copyright restrictions and the need for provenance, documentation and long-term control over data. Particular attention is given to independent legal opinions commissioned as part of KB’s government assignment, which conclude that Section 15 b of the Swedish Copyright Act provides a legal basis for developing language models through text and data mining for scientific research. Bio: Christopher Natzén is a Research Strategist at the National Library of Sweden (KB), which serves as both Sweden’s national library and a national research infrastructure. His role involves broad, overarching responsibility for the research-related matters within the National Library of Sweden’s remit, including research infrastructure and the strategic development of research activities. He is a member of the Swedish Research Council’s Advisory Group for Research Infrastructure of National Interest and serves as Sweden’s Scientific Representative in CESSDA ERIC. |
|
|
The Weight of Scarce Data: Perspectives from the Faroe Islands Iben Nyholm Debess, University of the Faroe Islands Abstract: From a Faroese perspective, data scarcity is not just about volume, but also access, copyright and quality. When resources are few, each dataset carries greater weight, raising questions of linguistic authority and language policy. Bio: Iben Nyholm Debess is a researcher in language technology at the University of the Faroe Islands. Alongside her PhD on evaluation and benchmarking for Faroese, she contributes to projects on machine translation, cultural representation in LLMs, and language data collection. |
|
|
|
|
|
National library data and LLMs in Finland – experiences and current situation Johanna Lilja, The National Library of Finland Abstract: The paper discusses opportunities for providing the National Library’s data for research that involves the use and development of AI‑based methods. It explains how the Tutkain and Fin‑CLARIN agreements regulate text and data mining in research, reflecting on the National Library’s experiences in delivering data. Finally, it reviews current debates on AI and copyright in Finland and asks how the research and GLAM sectors should promote opportunities for text and data mining, the use of AI, and the development of large language models (LLMs). Bio: Johanna Lilja is director of services in the National Library of Finland, and responsible for Collection and Data Services of the library. She is Doctor of Philosophy (Information studies, University of Tampere). Her previous academic work focuses on the history of learned publishing and her current professional papers on digitalisation of cultural heritage and copyright with regard using library materials. She leads WP data ingestion and engagement of cultural heritage sectors in the national research infrastructure project FIN-CLARIAH for SSH. |
|
|
Iceland and AI sovereignty Atli Þór Sigurgeirsson, Government of Iceland Abstract: As a small country, Iceland is in many ways an end user of AI technology, rather than an owner. This position, i.e. lack of AI sovereignty, enforces an external dependence which yields national security vulnerabilities. This talk motivates and defines AI sovereignty in the Icelandic context and covers the main initiatives the government of Iceland has implemented to increase the nation’s AI sovereignty. Bio: Atli Sigurgeirsson is a senior advisor on AI to the government of Iceland. He holds a PhD in natural language processing from the University of Edinburgh. His doctoral research focused on controllability for expressive Text-To-Speech. |
|
|
Connecting Academic Curation to AI Scale: The Language Bank of Finland in European Research and Digital Infrastructures Krister Lindén, University of Helsinki Abstract: For statistically smaller languages like Finnish and Finland-Swedish, the rapid advancement of generative AI poses a dual challenge: the scarcity of high-quality local training data and the aggressive commercial bias of English-dominated global models. To counter the risk of technological and linguistic domain loss, national language banks must evolve from traditional static repositories into active pipelines for large-scale language technology. This presentation outlines how the Language Bank of Finland (Kielipankki) leverages its 30-year legacy of research-driven data curation to anchor a sovereign Nordic and European AI ecosystem. We present a networked data-flow architecture where Kielipankki acts as the trusted national broker—clearing, annotating, and structurally refining high-value corpora. We examine how this local precision is federated for international academic research via CLARIN and DARIAH (through the FIN-CLARIAH framework), and subsequently scaled up to meet industrial machine-learning demands. Crucially, the presentation explores the emerging role of the Alliance for Language Technologies (ALT-EDIC) and the shared European Language Data Space (LDS). We demonstrate how curated academic assets can be safely ingested to fuel the training and benchmarking of localized, open-source foundational models (such as the Viking series or FinGPT) utilizing high-performance computing resources like the LUMI supercomputer. Finally, we advocate for deepened cross-border Nordic collaboration, showing that a synchronized infrastructure lifecycle between local curation, academic federation, and European industrial scaling is the only viable path to guaranteeing strategic digital autonomy for our national languages. Bio: Krister Lindén is a Research Director in Language Technology at the Department of Digital Humanities, University of Helsinki. He is a central figure in Finnish and European research infrastructures, serving as the Director of the Language Bank of Finland (Kielipankki), the National Coordinator for FIN-CLARIN, and he was the first Principal Investigator (PI) of the national FIN-CLARIAH consortium. On the European level, he is the Chair of CLARIN’s National Coordinators Forum (NCF) and a member of CLARIN’s Committee for Legal and Ethical Issues (CLIC). His research spans natural language processing (NLP), large-scale digital resource management, and computational methods applied to historical and linguistic data. |
Till konferenssidan | Conference page
|
08:30 – 09:00 |
Registration |
|
09:00 – 12:00 |
Session 3: The user perspective: from model to use |
|
Keynote: The Evolution of Evaluation Dan Smart, EuroEval, syv.aiAbstract: The types of models we’ve been using in NLP have changed many times over the course of the last decade, and every time we change, evaluation has to follow suit. I will be talking about the journey that evaluation methods has gone through and what’s ahead, while drawing parallels into the continuous development of EuroEval. Bio: Dan works as a Senior AI Engineer at the Danish startup syv.ai, has a PhD in Mathematics and have been working for 8 years within NLP in academia, the public and private sector. Dan is the creator and maintainer of the European LLM evaluation framework EuroEval. |
|
|
Coffee Break |
|
|
A Great Model Is Only the Beginning Tiina Lindh-Knuutila, LingsoftAbstract: State-of-the-art open models are a powerful starting point, but turning them into fit-for-purpose solutions requires more. This includes suitable licensing for commercial use, language and domain resources, adaptation, workflow integration and quality assurance. Using examples from professional language workflows, the presentation explores what is needed to turn model output into reliable results, and where human expertise fits into the process. Bio: Tiina Lindh-Knuutila, D.Sc. (Tech.), is Data Lead and a senior language technology expert at Lingsoft. She works at the intersection of research and practical applications. Previously a Solution Architect, she focused on applying models and technologies to real-world needs. She now leads Lingsoft’s data work, bringing together language resources, technology and practical use cases. |
|
|
Big data – what works and what doesn’t Tino Didriksen, Oqaasileriffik (The Language Secretariat of Greenland)Abstract: Benefits and shortcomings of LLMs trained on Greenlandic from the wider internet. Bio: Tino is a software engineer for Oqaasileriffik (The Language Secretariat of Greenland), implementing tools for linguists and end-users. |
|
|
From anecdotes to knowledge on AI language quality Matias Jentoft, The Language Council of NorwayAbstract: How do we make sure our national Nordic languages survive the arrival of groundbreaking technology? The technology has to be available, people need to use it, and the quality of language has to be good. The last few years LCN has been evaluating the quality of language in various applications, and publishing the results. Our final goal is to be able to give up-to-date advice to the public sector and others on how to deal with the situation. But we already have benchmarks, so why do we need human testing? In this talk I will describe how the fast pace of AI development and rollout has forced us to do human evaluations to contribute to the national discourse. We have tested chatbots, spellcheckers, and machine translation. I will present the results and our self-developed methodology. Bio: Matias Jentoft is working in the Language Council of Norway (LCN), on language policies in the context of technology. Important topics in focus are language coverage, language quality, and digital sovereignty. He finished his masters of Language Technology at the university in Oslo in 2023, with a thesis on grammatical error correction in Norwegian. He has been a proponent of Nordic cooperation since birth. |
|
|
12:00 – 13:00 |
Lunch |
|
13:00 – 16:00 |
Session 4: Nordic collaboration and strategic autonomy |
|
New Nordics AI Network and the Nordic LLM Network – Presenting a new collaboration framework for Nordic LLM Builders Jussi Karlgren, Nordic LLM NetworkAbstract: Det nordiska ministerrådet finansierar projektet ”Nordiskt nätverk för språkmodeller” som ett första steg i att skapa samarbetsformer för nordiskt språkmodellarbete, för att identifiera gemensamma utmaningar, för att dela resurser, teknik och metodologi. Projektet härbärgeras av New Nordics AI Network, ett nyinrättat samarbete mellan de nordiska och baltiska ländernas AI-noder. Denna presentation berättar om projektets aktiviteter och planer framöver. |
|
|
AI sovereignty, multilinguality, cultural fit, instruction training, and evaluation or How To Create Sustainable Language Resources Speaker TBD, Nordic LLM NetworkAbstract: As one of its first activities, the project ”Nordiskt nätverk för språkmodeller” has on Monday Nov 9 organised a brainstorm workshop to discuss requirements for training data for tomorrow’s language model training, with special interest for languages in the Nordics, both majority and minority languages. The objective of the workshop is to identify how memory institutions, language care organisations, media houses and publishers, and language communities best can devote resources to aid future generations of language models. Firstly: what types of training data are most crucial and helpful for model development efforts, how they should be shared and distributed, and what licensing conditions would be most appropriate. Secondly: how quality criteria for language models should be formulated. The presentation will be given by the participants from the workshop. |
|
|
Practical exercise for language model creation KortteliteatteriAbstract: To illustrate communicative cultural differences and how instruction training data may reflect them, performers from the Kortteliteatteri ensemble will enact some interactions between user and generative AI. The session will be interactive, and the audience will be able to contribute to the session! |
|
|
Conclusions and paths forward: What language resources would we need to create language models for all languages in the Nordics? Requirements and expectations from the technology developers Open discussion led by the Nordic LLM NetworkIntroduction: Given the findings from the Requirements brainstorm workshop and the preceding illustrative enactment, the strands will be brought together in a conversation about paths forward, and contributions from the present sectors. The objective is to put together a white paper on recommended methods and collaboration forms for open source technology developers, commercial developers and stakeholders, academic researchers, memory institutions, downstream users organisations, and language communities. |
Last modified on 2026-11-07