An Email Assistant That Learns to Suggest Reusable Replies

Abstract

Reusing previously authored email message text in new messages is an especially useful way to reduce workload for people such as IT administrators and event coordinators who tend to make repetitive and complex replies. Previous approaches to helping users find reusable message text have emphasized automation, minimizing the need for user interaction but requiring extensive training on examples of text reuse before achieving high performance. Our work focuses on extending this kind of assistance to handle ”spiky” patterns of reuse in which the user receives numerous similar queries over a limited time window, possibly related to some transient event. In this case, the system needs to learn rapidly to be useful. We have implemented an approach in which rapid improvement in performance at suggesting reusable text is achieved by learning from limited user feedback. The paper describes the implemented system, learning approach and an experiment to assess the impact of learning with a small number of training examples.

Keywords: Email, Information Retrieval, Machine Learning