Graphs and Spatial Relations in the Generation of Referring Expressions
Jette Viethen, Margaret A. Mitchell, Emiel Krahmer · 2013
When they introduced the Graph-Based Algorithm (GBA) for referring expression generation, Krahmer et al. (2003) flaunted the natural way in which it deals with relations between objects; but this feature has never been tested empirically. We fill this gap in this paper, exploring referring expression generation from the perspective of the GBA and focusing in particular on generating human-like expressions in visual scenes with spatial relations. We compare the original GBA against a variant that we introduce to better reflect human reference, and find that although the original GBA performs reasonably well, our new algorithm offers an even better match to human data (77.91 % Dice). Further, it can be extended to capture speaker variation, reaching an 82.83 % Dice overlap with human-produced expressions.