Graph-to-Graph Meaning Representation Transformations for Human-Robot Dialogue

Abstract

In support of two-way human-robot communication, we leverage Abstract Meaning Representation (AMR) to capture the core semantic content of natural language search and navigation instructions. In order to effectively map AMR to a constrained robot action specification, we develop a set of in-domain, task-specific AMR graphs augmented with speech act and tense and aspect information not found in the original AMR. This paper presents our efforts and results in transforming AMR graphs into our in-domain graphs by employing both rule-based and classifier-based methods, thereby bridging the gap from entirely unconstrained natural language input to a fixed set of robot actions.