Introduction

We now live in an ``age of information'', and arguably, communications serve as its motive force [3]. As participants in this age, we rely on cheap, ubiquitous, and reliable computers and communications to forward goals in our daily lives. Digital devices such as PDAs, laptops, and smartphones serve as our interface to this information-rich, electronic world.

In this world, many people have experienced typing an email on a smartphone keyboard such as the one depicted in Figure 1. A small keyboard and context-insensitive typing corrections lead to errors and a poor email experience. These devices require users to conform to their interfaces rather than the converse.

Figure 1: Email on the iPhone. Usabliity lacks despite the keyboard's QWERTY layout and email application's spelling suggestions. The suggestions lack context to prior emails and in addition, typing small letters with large fingers creates frustration. Some smartphones have tactile keyboards which do not fix these usability problems.

In this work, we will address the problem of unusable email by creating a machine learning (ML) model that accurately predicts ``the next word'' as a user types her email. We hypothesize that word prediction will provide a significant reduction in keystrokes for the following reasons: users have consistency in writing style, topics, and the identity of message recipients; a properly chosen learning algorithm can capture these similarities; and many users have a significant email corpus on which a learning algorithm can be trained.

To make email usable on a smartphone, we plan to use the following approach:

• We will focus our efforts on designing a suitable prediction model, as opposed to, for example, intregrating the model with an existing email application.
• As we build and refine the model, we will use a fraction of our personal email accounts as training datasets. This approach mirrors how one would use our model in practice.
• Finally, we will analyze its performance against the remaining fraction of personal email left out during training.

In the rest of this paper, we present more details of our proposal. We outine our approach in Section 2, describe the source of our training and testing data sets in Section 3, and present a rough project schedule in Section 4.