Keynote speakers
Arianna Bisazza is Associate Professor at the Center for Language and Cognition at the University of Groningen. She previously worked as a postdoc at the University of Amsterdam and got her PhD from Fondazione Bruno Kessler (Italy).
Her research revolves around two main directions: First, designing robust NLP algorithms that are equally capable of processing the large diversity of languages spoken around the world. And second, using neural networks to simulate processes of human language learning and the emergence of language universals. She has a long track record of contributions to machine translation for challenging language pairs, cross-lingual model adaptation, and neural model interpretability.
Recently, she was awarded a personal grant from the Dutch Research Council Talent Programme (VIDI) to work on child language acquisition-inspired LMs for morphologically rich languages.
Not all Language Models need to be Large: Studying Language Evolution and Acquisition with Modern Neural Networks
Abstract: Why do languages look the way they do? And what makes us so good at learning language as we grow up? Since the early days of connectionism, outstanding questions about human language have been investigated by means of simulations involving small neural networks (NNs) and toy languages. Is this still possible and meaningful in the age of Large pre-trained Language Models (LLMs)?
In this talk, I'll propose that modern NNs can indeed be a valuable tool to simulate and study processes of language evolution and acquisition. This, however, requires having control of training data, model architecture, and learning setup, which is typically not possible with LLMs.
I will then present two lines of research following these principles, namely: (1) simulating language change using small NN-agents that learn to communicate with pre-defined artificial languages, and (2) simulating the acquisition of syntax by training LMs on child-directed language. I'll end with a discussion of the value of interdisciplinarity and the importance of experimenting in controlled setups, rather than focusing all our research efforts on the evaluation of LLMs.
Keynote-TALK-ABisazza.pdf
Dirk Hovy is a Professor in the Computing Sciences Department of Bocconi University in Milan, and the scientific director of the Data and Marketing Insights research unit. Previously, he was faculty at the University of Copenhagen, got a PhD from USC’s Information Sciences Institute, and a sociolinguistics master’s from the University of Marburg in Germany.
He is interested in what computers can tell us about language and what language can tell us about society. Dirk is also interested in ethical questions of bias and algorithmic fairness in machine learning. He has authored over 100 articles on these topics, including 4 best and outstanding paper awards, and two textbooks on NLP in Python. Dirk has co-founded and organized several workshops (on computational social science and ethics in NLP), and was a local organizer for the EMNLP 2017 conference in Copenhagen. He was awarded an ERC Starting Grant project 2020 for research on demographic bias in NLP.
Outside of work, he enjoys cooking, leathercrafting, and picking up heavy things to put them back down.
The Illusion of Understanding – Unpacking the True Capabilities of Language Models
Abstract: The rapid development of large language models in recent years has transformed the field of NLP. Many people are concerned that it has trivialized the field or even rendered it obsolete.
In this talk, I'll argue that neither is true: NLP has a long way to go, and LLMs are the most recent in a long line of methods that have advanced the field. LLMs have freed us from many of the nitty-gritty details that previously hampered NLP research, allowing us to focus on larger and more interesting questions.
One of the most fundamental questions is what it means to "understand" language. In a world where AI can generate anything from translations to poetry and code, it's easy to believe these models genuinely understand us. However, despite its linguistic abilities, today's generative AI still resembles a skilled mimic rather than a genuine linguist. We will look at thought experiments and real-world examples to demonstrate the limitations of statistical models' knowledge, their inability to grasp context and nuance, and the dangers of overestimating their abilities. I will emphasize the theoretical and practical implications for future language technology, with a focus on social context. Drawing on philosophy, linguistics, and NLP history, we will investigate what it truly means to 'understand' a language beyond the words and the implications for safety and utility in LLMs.
slides_Dirk-Hovy.pdf
Arvi Tavast is director at the Institute of the Estonian Language, previously faculty at University of Tartu and Tallinn University.
After growing up at the Institute of Cybernetics and graduating from Tallinn University of Technology as a systems engineer, he drifted into translation, software localisation and IT terminology. Puzzled by how easy it was to falsify the structuralist and generativist ideas of language held by fellow practitioners, he did his MA and PhD in linguistics at University of Tartu. The first satisfactory answer, however, to why we then understand each other if not by encoding and decoding, he found from the intersection of psychology and information theory during his postdoc at University of Tübingen. He has authored papers on dictionary data models and intellectual property rights for language resources, as well as a university textbook on multilingual specialised communication.
Arvi’s hobbies cluster around folk music and endurance sports; when lifting, he prefers lightweight people to heavy things.
No Sex, No Future: On the Status of Estonian in a Changing World
Abstract: Apart from well-known anecdotes about the absence of gender marking and future tense, the most peculiar feature of Estonian is its number of speakers. Being one of the smallest fully functional languages in the world, it is a source of pride for its speakers, as well as a central part of their identity. The resulting puristic attitudes towards language also enjoy strong legal support. One of the enablers of this ideological stance is the channel metaphor of communication: that language as a system exists independently of its speakers, and communication works in virtue of using a shared code to encode and decode messages. This metaphor is still going strong in folk linguistics despite all evidence to the contrary, including recent advances in language modelling. A completely different reading for the title of the talk is provided by more recent learning- and prediction-based accounts of why we understand each other. Language, like any naturally evolving system, is vitally dependent on the random variability that is so conveniently present in linguistic data. This makes openness to new information a precondition to having a future, also for languages.