LIGHT, a Multi-Player Crowdsourced Text Adventure Game for Dialogue Research – I worked with my team to build a multiplayer crowdsourced text adventure game called LIGHT. We’ve made the complete setup open source and available to other researchers. Much current research today focuses on statistical regularities of fixed language data without an explicit understanding of the world that that language is trying to describe.
So for this game we wanted to build it entirely on natural language and as such, all of the locations, objects and characters within were written by people. Agents in LIGHT are trained on data produced by people playing the game as well which means that all of the environment communication and actions are natural and rich.
For example, the language present inherits all of the complex properties of natural language such as ambiguity and co-reference which can be challenging for machine learning models to understand but are an important part of the language that we use. I’ll talk a little bit about how we made the AI-controlled agents work.
First we devised an AI model that could produce separate representations for all of the contexts including the setting, a character’s persona and objects present in the room. And then we create a context embedding to score the most promising text candidates.
We used our PyTorch machine learning framework and ParlAI to build this model, then we use the Bidirectional Encoder Representations from Transformers, otherwise known as BERT model, that’s able to access context from both past and future directions to build two systems, a bi-ranker model which is fast and practical and a cross-ranker which is a slower model that allows more cross-correlation between the contexts we’ve calculated and the response and lastly, we used another set of AI models to encode the context and dialogue into features that let us generate game actions.
Combining these models we have characters in the game that can communicate with players and take relevant game actions. Before LIGHT we didn’t have a platform for studying language and actions jointly in a rich game world. Now that we’ve built that and made it openly available and showed its potential, we hope to enable future research in improving the ability of agents to think, speak and act, all while modeling a holistic world including the other agents within it.
Web enthusiast. Thinker. Evil coffeeaholic. Food specialist. Reader. Twitter fanatic. Music maven. AI and Machine Learning!