Resolving References in Natural Language Explanation Requests about Robot Behavior in HRI

Florian Schröder, Sonja Stange, Stefan Kopp

Abstract: In HRI, users have been shown to request explanations when their interpretation of autonomous robot behavior fails. These requests can refer to the behavior either by open questions or with attributes of the behavior. The presented work aims to resolve these references in explanation requests by developing an episodic memory with a graph database that stores and queries representations of the internal execution. The reference resolution is done by the detection of temporal adverb and verb constraints in the syntactical dependency tree of utterances, the execution of a query in the episodic memory, and the scoring of the resulting entries to find the referred behavior. The explanation generation process of the original model is adapted to the new approach and can contain additional information such as detected constraints, a failed execution state, and the distinction between running and completed executions.