Detecting comparatives in images

Abstract

Predicting relations such as special relations and verbs has been given importance in the area of Image understanding, but the area related to the detection of the comparative relation has been relatively unexplored. This work provides a dataset consisting of images with identified subject-object pairs, and a comparative between them. In this task, we also explore the effectiveness of different models in predicting the comparative based on language and geometric features from the subject and object bounding boxes. A discussion evaluating the results obtained from the designed models, as well as the relation of predicted comparatives to the synonyms of the gold label was also carried out.