Learning to Speak and Act in a Fantasy Text Adventure Game

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 673–683, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 673 Learning to Speak and Act in a Fantasy Text Adventure Game Jack Urbanek1 Angela Fan1,2 Siddharth Karamcheti1 Saachi Jain1 Samuel Humeau1 Emily Dinan1 Tim Rockt¨aschel1,3 Douwe Kiela1 Arthur Szlam1 Jason Weston1 1Facebook AI Research 2LORIA, Nancy 3University College London light-dms@fb.com Abstract We introduce a large-scale crowdsourced text adventure game as a research platform for studying grounded dialogue. In it, agents can perceive, emote, and act whilst conducting di- alogue with other agents. Models and humans can both act as characters within the game. We describe the results of training state-of-the-art generative and retrieval models in this setting. We show that in addition to using past dia- logue, these models are able to effectively use the state of the underlying world to condition their predictions. In particular, we show that grounding on the details of the local environ- ment, including location descriptions, and the objects (and their affordances) and characters (and their previous actions) present within it allows better predictions of agent behavior and dialogue. We analyze the ingredients neces- sary for successful grounding in this setting, and how each of these factors relate to agents that can talk and act successfully. 1 Introduction There has been remarkable progress in language modeling (Jozefowicz et al., 2016; Devlin et al., 2018; Radford et al., 2019) and building dialogue agents (Dinan et al., 2019a). Nevertheless, the current state of the art uses only the statistical regularities of language data, without explicit un- derstanding of the world that the language de- scribes. This work is built on the hypothesis that dialogue agents embodied in a rich and cohesive (but tractable) world can more easily be trained to use language effectively than those only exposed to standard large-scale text-only corpora. To that end, we introduce the LIGHT1 research platform. LIGHT is a multi-player fantasy text ad- venture world designed for studying situated di- alogue, and allows interactions between humans, 1 Learning in Interactive Games with Humans and Text. models as situated agents, and the world itself. It consists of a large crowdsourced game world (663 locations, 3462 objects and 1755 characters) de- scribed entirely in natural language. Within that game world, we collect a large set (11k episodes) of character-driven human-human crowdworker interactions involving actions, emotes, and dia- logue, with the aim of training models to engage humans in a similar fashion. Our complete frame- work is made publicly available in ParlAI (http: //parl.ai/projects/light). We use the collected dataset to investigate how a model can both speak and act grounded in per- ception of its environment and dialogue from other speakers. This is done by evaluating state-of-the- art models on our task and evaluating the effects of providing additional grounding. In particular, we adapt the BERT contextual language model (Devlin et al., 2018) to the task of dialogue in …

Content truncated to 500 words