Episodic memory for Companion Dialogue

Abstract

Artificial companions are interfaces designed to collaborate with a user over an extended period of time and communicate using natural language. This creates special requirements on the dialogue capabilities of companions. For acceptable long-term interactions, companions need to be aware of their own history and past interactions with their individual users. This can be accomplished by integrating episodic memory - the memory that remembers personal experiences - into the agent's cognitive model and dialogue capabilities. The thesis investigates on the one hand the design and implementation of an episodic memory for such artificial companions. On the other hand, how episodic memory can be used to enhance the dialogue capabilities of a companion. Based on a review of related literature, the required components are modeled and implemented. The model presented in the thesis thus takes into account the data structures and semantic representations of companions, and stores memories as RDF representations of the agent's experiences that are connected with semantic memory and a model of the interests and preferences of the user. The dialogue manager developed for the thesis uses a combination of rule-based processing and a scoring approach on a set of patterns operating on memories of the agent, the user's preferences, and the current input. The result of the thesis comprises two parts. First, the modelling and implementation of the respective components in the Java programming language using the Sesame RDF framework and the Elmo toolkit. Second, a set of use cases show application scenarios where the memory and dialogue manager enable the companion to refer to and comment on past interactions with the user, and to generate output taking into account the contents of the memory, the preferences of the user, and the current situation.