Tackling AlfWorld with Action Attention and Common Sense from Language ModelsDownload PDF

28 Sept 2022, 09:46 (modified: 16 Nov 2022, 23:48)LaReL 2022Readers: Everyone
Keywords: Language Models, Zero shot, AlfWorld, text games
TL;DR: We present (1) a novel question answering framework to simplify observation and (2) an action attention framework to handle large and variable size action space for the AlfWorld simulator.
Abstract: Pre-trained language models (LMs) capture strong prior knowledge about the world. This common sense knowledge can be used in control tasks. However, directly generating actions from LMs may result in a reasonable narrative, but not executable by a low level agent. We propose to instead use the knowledge in LMs to simplify the control problem, and assist the low-level actor training. We implement a novel question answering framework to simplify observations and an agent that handles arbitrary roll-out length and action space size based on action attention. On the Alfworld benchmark for indoor instruction following, we achieve a significantly higher success rate (50% over the baseline) with our novel object masking - action attention method.
3 Replies