Seminari del CoLing Lab – Dott.ssa Aurelie Herbelot, 29 maggio

    Lunedì 29 maggio alle ore 11 presso l’Aula Magna di Palazzo Matteucci ( Torricelli 2), la Dott.ssa Aurelie Herbelot (Università di Trento) terrà un seminario dal titolo:

    Ideal words: how to refer like a god

    Abstract: For a speaker to produce a satisfying reference, an appropriate referring expression must be generated, which uniquely identifies a given entity in the world for the hearer. Explaining this phenomenon with model-theoretic semantics involves a set of unreasonable assumptions with regard to speakers’ common ground: in effect, the standard approaches hypothesise omniscient individuals with a fully shared lexicon and semantic model. In return for this idealisation, however, they gain perfect ability to refer. In this talk, I will show that this idealised account of reference may be a good starting point to formalise what actually goes on between real speakers. I will focus on relaxing two assumptions in the ideal setup: the shared model and the shared symbols. This will require a new formalisation of the model and lexicon, as well as a revised denotation function. To achieve this, vectorial representations of sets and words will be used. In this modified framework, I will illustrate how the need for satisfying reference pushes speakers to approximate the ideal setup, and how the vector-based part of the semantics affords soft inference tools to build such approximation.

    Dr. Aurelie Herbelot obtained a PhD in Natural Language Processing from the University of Cambridge. Subsequently, she was an Alexander von Humboldt Fellow in Potsdam, and a postdoc in Cambridge and Stuttgart. She currently works at the Centre for Mind/Brain Sciences in Trento and will soon join the Universitat Pompeu Fabra in Barcelona as a Marie Curie fellow. She is particularly interested in models of semantics that bridge across formal and distributional frameworks, as well as vision and language models of quantification. In her spare time, she manages an open source project called PeARS, which aims at building a distributed and linguistically-informed search engine.

    Pubblicato il 25 Maggio 2017