Boleda Torrent, Gemma
Tenure-track researcher (Ramón y Cajal fellow)
Engineering Sciences
Short biography
Research interests
I want to understand how language works and be able to build machines that talk. In particular, I am interested in how talking agents convey meaning through language. I address this research question from a cross- and interdisciplinary perspective rooted in Computational Linguistics / Natural Language Processing and including Artificial Intelligence, Linguistics, and Cognitive Science.
In the past five years I have contributed to two major approaches to meaning, based on symbolic and distributed representations, respectively. Distributed representations, in particular those of distributional semantics and neural networks, are abstractions over language use in large amounts of data. Distributed models excel at capturing generic or conceptual knowledge ("Dogs can sleep"), but they have problems with episodic information ("My dog Lola is sleeping"), because they are very data-hungry and there is inevitably less data for instances (my dog Lola) than for concepts (dogs in general). Symbolic models, using discrete, logic-based representations, have opposite strengths and weaknesses. I have pushed for the integration of the two traditions in my own research and co-editing a Special Issue of the journal "Computational Linguistics".
In my current research, funded by an ERC Starting Grant, I am concerned with reference, or how we use language to talk about the world, and I am focusing on distributed models, while still keeping crucial insights from the symbolic tradition. Our neural network models can learn to refer directly from examples; our main innovation is to use a module for entity representations that can use both generic information associated to words like "dog" and episodic information coming from the context of use ("is sleeping"). An initial version of this architecture has won an international comptetition on a referential task (SemEval 2018 Task 4), and we have shown that its insights extend to other linguistic tasks like language modeling (word prediction).
Key words
Computational Linguistics; Natural Language Processing; computational semantics; semantic theory; distributional semantics; neural networks / deep learning