In today's world most of the successful shops, manufacturers and service-providers have already recognized the importance of customer service, which plays a key role in client retention. Besides, real-time text-based customer engagement's popularity is on the rise among these companies. Maintaining such a service requires a great amount of resources, so the question arises, can we automate client interactions and make operators obsolete?
To generate automated responses it is essential to create high-level representations of dialogues. Dialogue act tagging is a task in which we try to assign acts to each message from a predefined label set, capturing their underlying intentions.
In this thesis I compare multiple supervised and semi-supervised machine learning algorithms with different feature representations, in order to find the best model suitable for detecting frequently observed acts in a specific company's customer service dialogues, therefore providing a high-level model for these types of conversations.