Siri is Apple's virtual assistant that handles billions of user requests every week. Siri's natural language understanding component understands when you ask about the weather, want to make a phone call or want to update your calendar. This talk will present recent work by the Siri NLU team in two areas of conversational AI. Firstly, we discuss the issue of label error in data annotation. How often is an annotation provided by human labellers incorrect? What happens when label error exceeds our measurement of model error? We discuss a method for reducing label error while minimising the increase in annotation cost. The second topic presents a new task of conversational semantic parsing. In this task, the system must understand complex user utterances while maintaining a representation of the user's intent over multiple turns. We describe a method for generating a dataset for this task, as well as an encoder-decoder architecture for conversational semantic parsing. We demonstrate the benefits of a compositional meaning representation over a flat meaning representation of the sort traditionally used in state tracking for dialog systems.