Investigating linguistic strategies for abstraction

Concepts allow us to make sense of the world by structuring information. Organizing concepts hierarchically aids in abstracting and generalizing from one object to a more generic class. In this talk, I will explore the relationship between language and abstraction. Specifically, I will first show that participants tend to use superordinate referring expressions to refer to groups of objects in an interactive experiment. Speakers strategically communicate concepts at a level of abstraction that facilitates the listener’s comprehension. Second, I will investigate the linguistic strategies used in a computational model that features two artificial neural network agents that develop an emerging language while communicating about concepts at different levels of abstraction. We will look at the linguistic strategies that the agents use to generalize to concepts at novel levels of abstraction and see that agents tend to use a compositional strategy when generalizing to very specific concepts and a more flexible strategy when generalizing to very generic concepts. Finally, I will compare different pragmatic linguistic strategies and we will see whether context-based reasoning or recursive pragmatic reasoning modeled with the Rational Speech Act framework leads to the emergence of a more efficient language.

https://site.unibo.it/abstraction/it