Gemma Boleda Torrent
Fri 05 Feb 2016, 11:00 - 12:30
Informatics Forum (IF-4.31/4.33)

If you have a question about this talk, please contact: Diana Dalla Costa (ddallac)


Distributional semantics is a very successful, radically empirical, scalable, and flexible approach to meaning. Its representations (including those obtained with deep learning models) account for generic properties that are akin to conceptual knowledge:

Cats are similar to dogs, semantically related to veterinaries, and not very plausible agents of flying. However, when it comes to referring to a particular cat with specific properties, distributional semantics doesn't fare very well --and yet, reference is crucial to language, since we use words to talk about things in the world. I will review some referential phenomena that a more comprehensive model of meaning should handle, and discuss some theoretical and empirical work towards modeling reference with distributional semantics. In particular, I will show that distributional representations encode some referential properties of real-world entities such as countries and cities, and discuss the potential and limitations of learning a mapping between conceptual (distributional) and referential (database) representations.


Gemma Boleda works as a post-doctoral researcher at the University of Trento (Italy) on her Marie Curie grant on reference with distributional semantics. Before that, she earned a PhD at Universitat Pompeu Fabra (Spain) and carried out post-doctoral research in four universities including Stuttgart University (Germany) and The University of Texas at Austin (USA). She is currently serving as Area Chair for ACL 2016, as Information Officer in the ACL SIGSEM Board, and as co-editor of a Special Issue of the Computational Linguistics journal on Formal Distributional Semantics. In her research, Dr. Boleda uses quantitative and computational methods to better understand the semantics of natural language.