| title |
|---|
Invited speakers |
Director of Research CNRS at ISIR -- Institut des Systèmes Intelligents et de Robotique, Sorbonne Université
Multi-forms Adaptation for Socially Interactive Agents
Interacting with others enhances learning. Getting feedback on results, being encouraged and motivated… all help the learning process. During interaction, participants adapt to each other to show affiliation, group belongings, or to support social bonding. Adaptation can take place at different levels, through verbal alignment, imitation, and conversational strategies. Social resonance can also serve as a marker of adaptation. Socially Interactive Agents SIAs are virtual agents with a human-like appearance, capable of communicating verbally and nonverbally with their human interlocutors. In this talk, I will present our latest works aimed at endowing an SIA with various adaptive capabilities when interacting with its partners. The adaptation mechanisms are learned from human-human interaction data and evaluated by experimental studies involving human-agent interaction.
Professor of Linguistics and Computer Science at the University of Pennsylvania
Why language learning is not probabilistic
It seems harmless, and certainly mathematically convenient, to treat language learning as acquiring a probabilistic distribution over a space of linguistic patterns. The goal is to find or approximate an optimal hypothesis with respect to the data. Such is the mainstream machine learning approach, and the so-called Evaluation Procedure in generative grammar can be viewed as a particular instantiation.
Despite having pursued it vigorously in my earlier work, I now believe this approach is wrong (and wrong-headed). On the one hand, language is not a zero sum game: even overwhelming presence of one linguistic form does not necessarily inhibit or penalize alternative forms. On the other, the grammar can be a partial function: there are inputs for which no output form is acceptable even though some will always be most highly valued in a probabilistic framework.
The alternative is a theory of learning that does not even try to optimize but only sastifice. The coverage of the data only needs to be good enough up to a point; failure to do so may just result in the memorization of the input—nothing in the cognitive system mandates generalization under all circumstances. I will review the psychological and computational studies of the Tolerance Principle, a parameter-free learning theory that also appears operative beyond the domain of language.
Professor of Experimental Pragmatics at the Section of Theoretical and Applied Linguistics at the University of Cambridge and Fellow at Trinity College, Cambridge
“Deficit”, “difficulty”, “difference”: perspectives into autistic people's pragmatic skills and their implications for research methodology
It is widely reported that autistic people face pervasive challenges with producing and understanding pragmatics, i.e. context-dependent aspects of language. These are often attributed to challenges with mentalising, i.e. the ability to attribute the correct beliefs and intentions to other people. In this talk I will select influential papers from the past three decades of research in autism and language, each of which reveal a radically different perspective on the architecture of the linguistic system and on what it means to face challenges with linguistic competence (in our case, pragmatics). I will conclude that the recent perspective of neurodiversity implies a radical re-think of how we define pragmatics and how we assess the acquisition and processing of pragmatic competence.
Senior Researcher at CLASP and Professor emeritus of Computational Linguistics, Department of Philosophy, Linguistics and Theory of Science, University of Gothenburg. Acting Head of CLASP.
Types in a Theory of Interactive Learning
In this talk I will present some CLASP research on using types in a theory of interactive learning. In the first part of the talk I will introduce the notion of type we have been using and how it relates to a general theory of action, including linguistic acts. In the second part of the talk I will present some work I have been doing with Staffan Larsson and Jonathan Ginzburg on how such a theory relates to communicative acts by prelinguistic children and how such communicative acts serve as a basis for the development of linguistic acts by children. In the third part of the talk, I will present some preliminary work with Staffan, Jonathan and Andy Lücking on how the kind of types we are using might relate to the approach to neural modelling that Chris Eliasmith and colleagues have been developing. While this work is still in the very early stages, the hope is that ultimately we can propose an explanatory account of interactive learning which is grounded in biologically plausible neural activity.