Architectural Mechanisms for Situated Natural Language Understanding in Uncertain and Open Worlds

Chai et al. present a greedy As natural language capable robots and other agents become algorithm which uses a subset of the Givenness Hierarchy more commonplace, the ability for these agents to understand to resolve a wide array of referential expressions, but this truly natural human speech is becoming increasingly approach operates under a closed-world assumption (Chai, important. What is more, these agents must be able to understand Prasov, and Qu 2006). Kollar, Tellex et al. present Generalized truly natural human speech in realistic scenarios, Grounding Graphs, which instantiate probabilistic in which an agent may not have full certainty in its knowledge graphical models based on the structure of incoming NL utterances, of its environment, and in which an agent may not have and use those models to resolve references (Tellex full knowledge of the entities contained in its environment.