This paper investigates whether a curiosity-based strategy could be beneficial to word learning. Children are active conversation partners and exert considerable influence over the topics that are discussed in conversation with their parents. As the choice of topics is likely to be intrinsically motivated, a formalization of curiosity is implemented in a word learning model. The model receives annotated Flickr30k Entities images as input, and is trained in two conditions. In the curious condition, the model chooses objects to talk about from the scene according to the curiosity mechanism, whereas in the random condition, the model receives randomly chosen objects as input. The goal of this study is to show how a curious, active choice of topics by a language learner improves word learning compared to random selection. Curiosity is found to make word learning faster, increase robustness, and lead to better accuracy.

Models of cross-situational word learning typically character-ize the learner as a passive observer, but a language learn-ing child can actively participate in verbal and non-verbalcommunication. We present a computational study of cross-situational word learning to investigate whether a curious wordlearner who actively influences linguistic input in each contexthas an advantage over a passive learner. Our computationalmodel learns to map words to objects in real images by self-supervision through simulating both word comprehension andproduction. We examine different curiosity measures as guid-ing input selection, and analyze the relative impact of eachmethod. Our results suggest that active learning leads to higheroverall performance, and a formulation of curiosity which re-lies both on subjective novelty and plasticity yields the bestperformance and learning stability.