Serial dependency : Is it a characteristic of human referring expression generation?
Henriette Anna Elisabeth Viethen, Robert Dale, Markus Guhe · Research portal (Tilburg University) · 2011
A key characteristic of many existing referring expression generation (REG) algorithms is serial dependency: attributes are selected for inclusion one at a time, and the decision to include each attribute is dependent on the discriminatory ability of the set of attributes that have already been selected so far. We use a machine learning approach to explore whether serial dependency is a characteristic of human referring expression generation. Our results show that models in which attributes are chosen in a serially dependent fashion does not perform better than one where their inclusion depends only on other factors. The results also suggest that the visual salience of an attribute might be more important than its discriminatory power.