Connectionist Modelling of Quantifiers

EPSRC Grant GR/S26569

 

 

Summary     Objectives & Workplan       Background         Results        Publications

 

 

Summary

Quantifers in English, like 'few', 'some' and 'many' give information about the numbers of objects being talked about. Yet the actual number of objects in a scene to be described is not a perfect predictor of the use of these terms. In this grant we showed that people's quick estimates of the number of objects in a visual scene is a better predictor of quantifier appropriateness than the actual number. In a series of experiments we asked people either to estimate the number of objects in a visual scene (when presented for a short time), or to rate the appropropiateness of quantifiers (e.g., 'There are many fish') to describe the same scenes. We found a correspondence between the number estimates and quantifier ratings: variables such as the spacing between objects, how they are grouped together, and how many other objects are present all affect both number judgements and quantifier ratings. We then used this data to build a computational model of quantifier use grounding in visual knowledge of the scenes being described. The model takes as input a visual scene, and gives a quantifier description as output. Using the model, we were able to show that learning how to use quantifirs is faster if the model has knowledge about number estimation first, and learning number judgements is faster if the model has been training to use quantifiers first. These results show that language and perception are intricately connected, and that decoupling them adversley affects learning.

People

Kenny Coventry (Principal investigator; now at University of Northumbria)

Angelo Cangelosi (Co-investigator)

Steve Newstead (Co-investigator)

Alison Bacon

Rohana Rajapakse

Davi Bugmann