Abstract Meaning Representation for HumanRobotDialogue

摘要

In this research, we begin to tackle the\nchallenge of natural language understanding\n(NLU) in the context of the development of\na robot dialogue system. We explore the adequacy\nof Abstract Meaning Representation\n(AMR) as a conduit for NLU. First, we consider\nthe feasibility of using existing AMR\nparsers for automatically creating meaning\nrepresentations for robot-directed transcribed\nspeech data. We evaluate the quality of output\nof two parsers on this data against a manually\nannotated gold-standard data set. Second,\nwe evaluate the semantic coverage and distinctions\nmade in AMR overall: how well does it\ncapture the meaning and distinctions needed\nin our collaborative human-robot dialogue domain?\nWe find that AMR has gaps that align\nwith linguistic information critical for effective\nhuman-robot collaboration in search and\nnavigation tasks, and we present task-specific\nmodifications to AMR to address the deficiencies.