Dialogue-AMR Parsing Pipeline

Author(s):

Mitchell Abrams, Claire Bonial, and Clare Voss

Abstract:

This research forms part of a larger project focused on natural language understanding (NLU) in the development of a two way human–robot dialogue system in the search and navigation domain. We leverage Abstract Meaning Representation (AMR) to capture and structure the semantic content of natural language instructions in a machine-readable, directed, a-cyclic graph. Two key challenges exist for NLU in this task: 1) how to effectively map AMR to a constrained robot-action specification within a particular domain and 2) how to preserve necessary elements for general understanding of human language with the goal that our robot may expand its capabilities beyond a single domain. To address these challenges, we establish a two-step NLU approach in which automatically obtained AMR graphs of the input language are converted into “Dialogue-AMR” graphs, which is a new version of AMR that is augmented with tense, aspect, and speech act information. Here, we detail both rule-based and classifier-based methods to transform AMR graphs into Dialogue-AMR graphs, thereby bridging the gap from unconstrained natural-language input to a fixed set of robot actions.