Grounded language interpretation of robotic commands through structured learning

The presence of robots in everyday life is increasing day by day at a growing pace. Industrial and working environments, health-care assistance in public or domestic areas can benefit from robots' services to accomplish manifold tasks that are difficult and annoying for humans. In such scenarios, Natural Language interactions, enabling collaboration and robot control, are meant to be situated, in the sense that both the user and the robot access and make reference to the environment. Contextual knowledge may thus play a key role in solving inherent ambiguities of grounded language as, for example, the prepositional phrase attachment. In this work, we present a linguistic pipeline for semantic processing of robotic commands, that combines discriminative structured learning, distributional semantics and contextual evidence extracted from the working environment. The final goal is to make the interpretation process of linguistic exchanges depending on physical, cognitive and language-dependent aspects. We present, formalize and discuss an adaptive Spoken Language Understanding chain for robotic commands, that explicitly depends on the operational context during both the learning and processing stages. The resulting framework allows to model heterogeneous information concerning the environment (e.g., positional information about the objects and their properties) and to inject it in the learning process. Empirical results demonstrate a significant contribution of such additional dimensions, achieving up to a 25% of relative error reduction with respect to a pipeline that only exploits linguistic evidence.