Gameplay data research gained popularity in the last decade across several disciplines (e.g. linguistic, pedagogy, psychology) as very precious way to analyse behavioural patterns and interactions among humans.
One of the main limitations currently faced in gameplay research is the lack of open data. Gameplay data available on Youtube or Twitch is not suitable for research, which needs instead non-staged, everyday practice of players. The main reason behind the absence of an open dataset of gameplay data is the lack of tools to simply anonymize this type of data, which typically contains visual and textual identifiers of players. The primary goal of this project is to develop a machine learning pipeline that will anonymize gameplay data while preserving game dynamics and other information critical for research. We will leverage recent transformer-based multimodal AI methods to anonymize textual chat data and visual elements and therefore eliminate any references to player identities.