Instant Messaging chat sessions are realtime text-based conversations which can be analyzed using dialogue-act models. We describe a statistical approach for modelling and detecting dialogue acts in Instant Messaging dialogue. This involved the collection of a small set of task-based dialogues and annotating them with a revised tag set. We then dealt with segmentation and synchronisation issues which do not arise in spoken dialogue. The model we developed combines naive Bayes and dialogue-act n-grams to obtain better than 80% accuracy in our tagging experiment. .
