Intelligent email: Aiding users with AI

User productivity and attention suffer from email overload. The human computer interaction community has designed new types of interfaces to facilitate email management, including email triage, activity management, search and organization. In this work we draw ideas from machine learning and natural language processing to introduce intelligent email and define it as intelligent systems for supporting email interfaces. Our interfaces are information driven, enabling users to make faster, smarter and less error prone decisions in processing email.;We develop intelligent email in several stages. First, we examine the common problem of the forgotten attachment by building an attachment prediction system, which can support different user interfaces for this problem. Next, we explore the task of email triage, the process of managing large amounts of email. We propose a reply management system supported by a reply predictor, automatically labeling messages that need a reply. To enable cross-user learning, we develop a shared set of deictic features with user specific extraction based on social network analysis of email. We then explore new representations for message content based on latent concept models. Next, we develop a system for email activity classification to support email activity management interfaces. Finally, we extend the popular tool of faceted browsing to email by developing automatic facet rankers to select the most useful facets for display to the user. A large scale evaluation and user survey demonstrates the effectiveness of intelligent email applications in real world settings.;We also consider new learning methods useful for intelligent email: Confidence-Weighted (CW) learning. CW learning is a family of online learning algorithms where online updates are confidence sensitive, favoring larger updates to rarer features. Incorporating language sensitivities improves performance on a number of NLP applications, including important learning settings for intelligent email. We consider how to scale learning in the Email setting to very large data environments through parallel training. To reduce the cost labeling email by users, we consider active learning. We show that CW learning improves standard margin-based active learning. Finally, we show how confidence sensitive parameter combinations can be used to perform cross-user and multi-domain learning.