The role of corpora for robots’ acquisition of linguistic structure through observation and natural language instructions

Abstract

When aiming at automatic linguistic structure learning, the developed algorithms highly depend on the data they can be trained on. We present several multimodal datasets employed for grounded language learning in artificial agents. Based on evidence on the close tying of motor action and language, we developed the Linguistic, Kinematic and Gaze information in task descriptions Corpus (LKG-Corpus) as a resource to (i) investigate fundamental questions concerning the relation between sensorimotor processes and linguistic structure, and to (ii) develop computational models for grounded language learning in robots. For embodied structure learning, we emphasize the importance of data sets which can be automatically interpreted by the robot and do not need prior knowledge about linguistic structure or actions.