Games are a fantastic domain for studying systems that predict human intentions, because the human's goals and actions are easily defined, and hard perceptual problems are out of the way. One project we've just started is using Bayesian reasoning over time to decide what a player's current goal is in an open-ended adventure or RPG game. This information could then either influence what content is presented to the player next, or serve as a handy reminder to the player when picking up an old save game.

Understanding intentions and language learning can both come in handy in story-driven adventure games, where the player can sometimes be more clever than the designer in solving puzzles. Interactive fiction makes a natural proving ground for language-driven intention recognition, before either adapting it to more advanced games (Sims 3 drew on interactive fiction in its character interactions) or real-world applications.

Another current project (Alex) is automatically converting Guitar Hero difficulty levels using statistical machine translation. Game content can form an interesting language-like "corpus" on which to apply statistical NLP and machine learning, and given the game industry's reliance on data-driven game content, methods for extrapolating from existing content could be useful in increasing the longevity and replayability of existing games. (Image from Frets on Fire website.)