Learning to Speak and Act in a Fantasy Text Adventure Game

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 dialogue 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 dialogue, 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 environment, 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 necessary for successful grounding in this setting, and how each of these factors relate to agents that can talk and act successfully.

ALOHA: Artificial Learning of Human Attributes for Dialogue Agents

“…The limited human evaluation size limits what can be inferred, but it indicates the problem is solved to the extent that ALOHA is able to slightly outperform humans on two folds and perform closely on another two folds. Even humans do not perform extremely well, demonstrating this task is more difficult than typical dialogue retrieval tasks (Urbanek et al 2019;Dinan et al 2019).…”

“…These are trained to produce LSRM-BERT and LSRM-Poly, respectively. (Dinan et al 2019;Urbanek et al 2019;Zhang et al 2018) choose 20 candidate responses, and for comparison purposes, we do the same.…”

“…Candidate response selection is similar to the procedure from previous work done on grounded dialogue agents (Zhang et al 2018;Urbanek et al 2019). Along with the ground truth response, we randomly sample 19 distractor responses from other characters from a uniform distribution of characters, and call this process uniform character sampling.…”

Interactive Fiction Games: A Colossal Adventure

“…Finally, although not a sequential decision making problem, Light (Urbanek et al 2019) is a crowdsourced dataset of text-adventure game dialogues. The authors demonstrate that supervised training of transformer-based models have the ability to generate contextually relevant dialog, actions, and emotes.…”

Crowdsourcing Complex Associations among Words by Means of A Game

“…The aspect of interestingness is particularly difficult to preserve if a game needs specific training [17] or it involves (potentially long) free-text entries produced by the user [18]. Text adventure games provide an example of attractive environments that can naturally lead to building useful language resources [19]. The use of Codenames as the platform to collect valuable lexical data is also given by the popularity of the game and its attractiveness for various players across languages.…”