Marko Robnik Šikonja: Large language models for cross-lingual transfer
Affiliation: University of Ljubljana
Abstract:
Currently, the most successful approach to natural language processing is based on large pretrained language models using the transformer architecture of neural networks. These are typically pretrained on huge text corpora on the tasks of predicting next tokens or masked tokens. While most existing models are predominantly monolingual, multilingual variants also exist and can help in cross-lingual transfer of knowledge and models. We will present a few types of large language models, focusing on cross-lingual transfer. We will show their strengths and weaknesses in text classification, summarization, and question answering.
Speaker short bio:
Marko Robnik-Šikonja is Professor of Computer Science and Informatics at the University of Ljubljana, Faculty of Computer and Information Science. His research interests span machine learning, data mining, natural language processing, explainable artificial intelligence, and application of data science techniques. His most notable scientific results concern deep learning, natural language analysis, feature evaluation, ensemble learning, predictive model explanation, information network analysis, and data generation. He is (co)author of more than 200 scientific publications that were cited more than 7,500 times. He is the author of several data mining software packages and language resources.
Malvina Nissim: Language Technology <preposition> Society
Affiliation: University of Groningen
Abstract:
The recognition of society’s role in language technology has become essential and cannot be overlooked. Still, plenty of research in Natural Language Processing does not explicitly account for such interplay. This evening lecture will zoom in on precisely this aspect. “Precisely” is an ambitious term, since the very definition of the relationship between language technology and society is subject to multiple interpretations, both in the context of scientific research as well as in connection with the general public, who currently is very much exposed to, interested in, and involved with language-based artificial intelligence tools. Through recent work I’ve carried out with my group, and through personal reflections, I will unpack this exciting relationship from different angles.
Speaker short bio:
Malvina Nissim holds a Chair in Computational Linguistics and Society at the University of Groningen, The Netherlands. Her research focuses both on language modelling aspects as well as on the impact that language technology has on society. She's thus also regularly involved in outreach activities, and is a member of the Ethics Committee of the international Association for Computational Linguistics (ACL). She graduated in Linguistics from the University of Pisa, and obtained her PhD in Linguistics from the University of Pavia. Before joining the University of Groningen, she was a tenured researcher at the University of Bologna (2006-2014), and a postdoc at the Institute for Cognitive Science and Technology in Rome (2006) and at the University of Edinburgh (2001-2005). She is the 2016 University of Groningen Lecturer of the Year.
Beniamino Accattoli: The Cost of the lambda Calculus and the Semantics of Sharing
Affiliation: Inria
Abstract:
The lambda calculus is an expressive mathematical formalism that elegantly captures the core of functional programming languages, while providing at the same time compact representations of intuitionistic logic proofs. The first part of the talk shall survey the recent advances in the study of reasonable cost models for the lambda calculus, that is, of time and space cost measures that are equivalent to those of Turing machines. In particular, it shall overview how understanding the role of sharing in the evaluation process is crucial for both time and space, but for opposite reasons. The second part of the talk shall show that extending the lambda calculus with first-class sharing is not a minor extension, as crucial semantic properties and results break, and new tools and richer theories need to be developed.
Speaker short bio:
Beniamino Accattoli obtained his PhD at 'La Sapienza' University in Rome in 2011. Since 2015, he has been a researcher at Inria & Ecole Polytechnique.
His work is mainly about the lambda calculus, using a combination of tools from functional programming, logic, rewriting theory, and complexity theory. His most relevant contribution, in collaboration with Ugo Dal Lago, is that the lambda calculus can be used as a reasonable model for computational complexity.
Rick Nouwen: The language, logic and computation of exaggeration
(Dick Oehrle Memorial Lecture)
Affiliation: Universiteit Utrecht
Abstract:
Classic accounts of linguistic information exchange are based on the assumption that the meaning of an uttered sentence helps us navigate the space of all the possible ways the world could be, where this space is seen as a simple partitioned set or as the event space of a probability distribution. Underlying these models is the assumption that the speaker of the utterance is being truthful. It is well-known, however, that we often deliberately use sentences that we do not believe to be true, without the intention to deceive. Examples include metaphor and jokes, but in this talk I will focus on exaggeration. We routinely use utterances that over- or understate things. For instance, when you are very busy and someone asks you to do something, you may state that you've got "a million things to do", or say sarcastically "sure, I've got nothing else to do". In this lecture, I will look at linguistic, semantic and pragmatic properties of such utterances and show how we can accommodate them in existing logical and computational models of linguistic communication.
Speaker short bio:
Rick Nouwen obtained his PhD at Utrecht University in 2003. Since then, he has held positions at the University of Edinburgh, the J.W. Goethe Universitaet Frankfurt am Main and the Massachusetts Institute of Technology. He is now an associate professor at Utrecht University, where he teaches linguistics and artificial intelligence and where he is affiliated to the Institute for Language Sciences.
His research is in the domain of linguistic semantics and pragmatics, specialising especially in the communication of quantity and degree. He has published journal articles, books and book chapters on, among other things, plurality, comparative constructions, intensification, exclamation and numeral quantification, using philosophical, logical, (cross)linguistic, and computational perspectives.
He was the managing editor of Journal of Semantics (Oxford University Press) between 2013 and 2021. Before that he held associate editor positions at both Journal of Semantics and Semantics & Pragmatics. Between 2010 and 2015 he was the chair of the Dutch Linguistic Society. He was the recipient of an ERC Consolidator grant (2013) and several large personal grants from the Dutch Research Council (2005, 2008, 2018).
Rick Nouwen has (co)lectured at ESSLLI twice (2006, 2019) and (co)organised two ESSLLI workshops (2007, 2009).