Unsupervised discovery of human behavior and dialogue patterns in data from an online game

Abstract

A content authoring bottleneck in AI, coupled with improving technology, has lead to increasing efforts in using large datasets to power Al systems directly. This idea is being used to create Al agents in video games, using logs of human-played games as the dataset. This new approach to AI brings its own challenges, particularly the need to annotate the datasets used. This thesis explores annotating the behavior in human-played games automatically, namely: how can we generate a list of events, with examples, describing the behavior in thousands of games. First dialogue is clustered semantically to simplify the game logs. Next, sequential pattern mining is used to find action-dialogue sequences that correspond to higher-level events. Finally, these sequences are grouped according to their event. The system can not yet replace human annotation, but the results are promising and can already help to significantly reduce the amount of human effort needed.