Automating Chat

• In Event: 063 - Section on Ethnomethodology and Conversation Analysis Paper Session (one-hour). Ethnomethodological Studies of Work: Challenges, Practices, and Successes

Abstract

In this paper I will describe efforts underway to automate service chat and the challenges this endeavor is coming up against. Even though there are many advances in linguistic understanding and machine learning, there are aspects of human-to-human interaction that remain hard to replicate.
The paper is based on empirical research in call centers in which some agents do Chat support. Interactions were video recorded, the agents were interviewed and the chat sessions analyzed.
The paper will also discuss more recent efforts to expand the automation to “speech,” which is conceived to be the adding of two additional components to the Chat system: a natural language generator, which can transform symbolic representations into audible sound, and a natural language understanding module that can transform audio of talk into symbolic representations. I will argue that this model is flawed, and that it cannot lead to a successful conversationalist.

Author

• Erik Vinkhuyzen, Palo Alto Research Center