Say What? Analyzing the Impact of an Entity-Level Model of Working Memory Forgetting on Referring Expression Generation

Rafael Sousa Silva and Tom Williams

International Symposium on Technological Advances in Human-Robot Interaction

Working Memory (WM) is a necessary component both for models of human cognition and human-inspired robot cognitive architectures. Yet it is unclear how different parameterizations for Working Memory models might impact robot cognition, especially robots’ ability to engage in natural, situated, language-based interactions. In this work, we evaluate an entity-level, feature-based Working Memory framework through an analysis of temporal decay and demonstrate with a set of case studies how different parameterizations for this WM dynamic have fundamentally different error modes in different interaction contexts. Specifically, we formulate rules that inform the selection of appropriate decay rate values to be used in contexts with different environment dynamics and dialogue dynamics. By formalizing and analyzing these parameterizations within a robot cognitive architecture, we are able to make key design recommendations for robot cognitive architectures