Working Memory Models for Referring Expression Generation in Robot Cognitive Architectures

Rafael Sousa Silva

Colorado School of Mines PhD Dissertations

Memory is an important component of cognition, and is the system responsible for encoding, storing, and retrieving information. Working memory is the memory subsystem that enables real-time cognition by maintaining short-term, task-relevant information and keeping it accessible to other cognitive processes for immediate use. Working memory is especially important for language processes, including language comprehension, language acquisition, and referring expression generation. Given the benefits that working memory has to human cognitive processes and language, this research implements working memory into robot cognitive architectures using an attribute-based perspective to facilitate the process of referring expression generation. This work borrows inspiration from models of human working memory to enable better robot dialogue through the DIARC and LARCH cognitive architectures, which are built to enable language-based human-robot interactions in situated contexts. This work proposes three different designs for how working memory can be organized in these architectures. First, at a global level, working memory is implemented in a human-like way as a unitary storage system. Second, at a consultant level, working memory is split into several storage buffers, each responsible for storing information about one particular type of entity, such as people, objects, and locations. Finally, at an entity level, each entity has a working memory buffer that stores activated information. At all levels, forgetting mechanisms play an essential role regarding how working memory information is updated. To parameterize such models appropriately, this work proposes rules that guide the selection of values for how long information remains in working memory storage and for the overall working memory capacity for each configuration. Finally, this work evaluates whether the proposed working memory configurations and parameters are well suited for informing referring expression generation through human-subjects experiments. The results provide empirical support for using working memory to generate better robot referring expressions. Specifically, our results suggest that parameterizing working memory to promote lexical entrainment, to avoid the use of outdated information, to avoid cache misses, and to allocate space efficiently can lead to the generation of better robot referring expressions.