Classifying Dialogue Acts in Multi-party Live Chats

We consider the task of classifying chat contributions by dialogue act in a multi-party setting. This extends the problem significantly over the 1-1 chat scenario due to the semiasynchronous and “entangled” nature of the contributions by chat participants. We experiment with a number of machine learning approaches, using different categories of features: lexical, contextual, structural, keyword and dialogue interaction information. For evaluation, we developed gold-standard data using online forums from the USA Library of Congress. We found that, for multi-party dialogues, features based on 1-gram and keywords produced best performance, while features exploiting structure and interaction did not perform as well as previously reported results over 1-to-1 chats.