Text-Room Adventure

In interactive fiction, players interact with unknown game worlds described in natural language. Most of the interactions rely on commonsense knowledge. Previous works have tried to use knowledge graphs to represent commonsense knowledge explicitly. However, none of existing works utilized rich commonsense knowledge outside the game world.

In this paper, we incorporate pre-trained language models into the deep Q-learning framework to bring in rich commonsense knowledge existing in external text corpus, and combine history actions into decision making process.

Further, we build a 2D simulator that can visualize the game process to implement case study and assist detailed performance analysis. Experimental results on the Jericho platform show that our method performs better than several strong baseline methods.