CENTAL Seminars

👤  Speaker: Simone Balloccu

💬  Titre: How expert is your "expert AI" in the real world?

📝  Abstract: Large-scale data annotation, training, and alignment bring the promise of "Expert AI". Depending on who you ask, this ranges from models capable of assisting domain experts in their work to superhuman AGI, which may replace you in the near future. Evidence of this "expertise" often comes from benchmarks, leaderboards, and sometimes even ad-hoc cases. In this talk, we take a closer (and perhaps, critical) look at this evaluation paradigm: how "expert" are these models when we apply them to really challenging real-world scenarios? We will cover three hard cases: irrelevant information, subjective preferences, and enterprise QA. Our "expert AI" might gracefully collapse, teaching us a lesson on what we need to do before we deploy these systems in sensitive domains.

Short bio: Simone Balloccu is a computer scientist with 9 years of research experience in NLP & AI. He worked within several EU-funded projects, including Horizon 2020, and ERC, focusing on AI for mental health and behaviour change, human evaluation, and more generally on AI applied to expert domains. He leads the “NLP for expert domains” lab at TU Darmstadt, researching Expert-AI interaction and cooperation. His current research involves efficient RAG systems over corporate knowledge, Multimodal NLP applied to mental health, and modelling expert preferences in LLMs.

📅  Date:  02/10/2026

📍  Local: Maison Des Langues A118 (FIAL)  ( Google Map link ) 

👤  Speaker: Els Lefever 

💬  Titre: From shared meaning to competing hypotheses: how large language models represent meaning, uncertainty, and alternative interpretations.

📝  Abstract: Large language models achieve remarkable performance across languages and reasoning tasks. Yet, high benchmark scores reveal little about how models internally represent information or how they reason when evidence is incomplete. To better understand these capabilities, this talk combines insights from mechanistic interpretability and defeasible reasoning.
First, I examine how multilingual language models represent meaning at the neuron level. Using cross-lingual emotion detection as a case study, I show that multilingual representations are neither fully shared nor fully language-specific. Mechanistic analyses reveal a dual structure consisting of polyglot neurons that encode language-agnostic semantic information and monolingual neurons that capture language-specific patterns.
I then turn from how information is represented to how it is used in reasoning. Defeasible reasoning provides a revealing test case because conclusions must remain open to revision when new information becomes available. Experiments with counter-hypotheses show that explicitly representing alternative interpretations improves reasoning, while question-guided prompting can surface both an original hypothesis and plausible alternatives.
Taken together, these studies argue for a shift from asking only whether LLMs get the right answer to asking how they arrive there. By examining which representations are shared, which are local, which alternatives are considered, and how uncertainty is encoded in the reasoning process, we gain a more nuanced understanding of the capabilities and limitations of current language models.

📅  Date:  13/11/2026

📍  Local: Maison Des Langues A118 (FIAL)  ( Google Map link ) 

👤  Speaker: Sylvain Kahane

💬  Titre: Grammaires et typologie grammaticale à partir de treebanks

📝  Abstract: Après une rapide présentation de la notion de corpus arboré ou treebank et un historique de leur développement depuis le 18e siècle au moins, nous intéressons à deux applications de ces ressources. Premièrement, l’utilisation des treebanks pour l’induction de grammaires, y compris des grammaires contrastives, avec notamment les projets ANR Autogramm et Autogramm IA. Et deuxièment, l’exploitation de la collection de treebanks Universal Dependencies avec plus de 180 langues pour des études typologiques. Nous intéresserons en particulier à la taille des constituants et une propriété qui veut que plus les compléments sont éloignés du verbe, plus ils sont longs.

📅  Date:  27/11/2026

📍  Local: tbd

👤  Speaker: Dirk Hovy

💬  Title: Unsolved: Why the Most Interesting Problems in NLP Are Still Ahead

📝  Abstract: Large language models now power everyday writing, search, tutoring, and decision support. It might seem that with all that, natural language processing is essentially solved. However, that ignores the nature of language. While LLMs are remarkably fluent and better than humans at information processing, scale and fluency are not the same as language understanding. Language is not just a vehicle for information transfer: it is shaped by context, norms, relationships, ideology, and communicative goals. In this talk, I will argue that we have solved one set of problems, only to be able to explore a new, even more interesting set of challenges along those dimensions.

Drawing on recent work on disagreement, moral judgment, safety evaluation, and socioeconomic differences in AI use, I will show that current models capture linguistic patterns without fully modeling the social aspects of language. I will argue that this is a hopeful moment for the field: we are finally in a position to address richer questions about social reasoning, human diversity, and language in interaction.

The central claim is that NLP is far from finished. Its next frontier is not simply better next-token prediction, but deeper engagement with the social nature of language and closer collaboration with the social sciences, opening exciting new research directions.

👤  Bio: Dirk Hovy is a professor in the Computing Sciences Department, the scientific director of the Data and Marketing Insights research unit, and the inaugural dean for Digital Transformation and AI of Bocconi University. Previously, he was faculty at the University of Copenhagen, got a PhD from USC’s Information Sciences Institute, and a linguistics master’s from Marburg university in Germany. 

Dirk is interested in what computers can tell us about language and what language can tell us about society. That involves ethical questions of bias and algorithmic fairness in AI. Dirk has authored over 150 articles on these topics, two textbooks on NLP in Python, and forthcoming title at MIT Press. Dirk has co-founded and organized several workshops (on computational social science, and ethics in NLP), was a local organizer for the EMNLP 2017 conference, and general chair of EMNLP 2025. He was awarded an ERC Starting Grant project 2020 for research on demographic bias in NLP. 

In his spare time, he enjoys cooking, leather working, and picking up heavy things to put them back down.

📅  Date:  04/12/2026

📍  Local: Maison Des Langues A118 (FIAL)  ( Google Map link ) 

👤  Speaker: Clara Degeneve

💬  Titre: Données textuelles et science forensique : L’exploitation des traces langagières pour l’analyse de phénomènes de criminalité en ligne

📝  Abstract: Depuis les années 50, le numérique ne cesse de prendre de plus en plus de place dans le quotidien. Ceci a entrainé non seulement un déplacement de la criminalité vers les espaces numériques, mais également le développement de nouvelles formes de criminalité propres à ces espaces.  De nouvelles problématiques ont également émergé pour les professionnels de la sécurité, qui doivent alors faire face à une masse de données et de traces beaucoup plus importante qu’auparavant. Parmi ces traces, celles issues des réseaux sociaux, des plateformes en ligne et des applications de messagerie instantanée sont nombreuses. L’exploitation des données textuelles dans un cadre judiciaire pose un certain nombre d’enjeux, notamment sur l’objectivité, la robustesse et la reproductibilité des méthodes issues de la linguistique computationnelle. A travers un exemple d’analyse des traces langagières liées à des cas de fraude en ligne, cette présentation tentera de mettre en lumière les forces et faiblesses de l’usage de la linguistique computationnelle dans un contexte d’analyse de phénomènes de criminalité numérique.

📅  Date:  11/12/2026

📍  Local: Maison Des Langues A118 (FIAL)  ( Google Map link )