Automatically detecting player roles in Among Us

Player role identification in the online social deduction game Werewolf has been a common research topic in the past decade. Among Us, a fairly new online social deduction game, gained a lot of popularity recently. Given the popularity of the game, an opportunity arises to extract information on social deduction. This research focuses on extracting information from emergency meetings in Among Us and subsequently using this information to automatically detect player roles. First, a framework is presented that can be used to extract information from videos with gameplay of Among Us. This framework can extract the chat messages from the emergency meetings, detect imposters at the end of the game, and extract voting data. Secondly, the framework is used to process videos with 59 games of Among Us gameplay. The data produced by the framework will be normalized and transformed into numerical data using term frequency-inverse document frequency (tf-idf), Finally, two types of Support Vector Machines (SVM) and a Naive Bayes classifier are used to automatically detect player roles. We show that detecting player roles is a difficult yet learnable challenge.