Introduction

In this work, we proposed to reduce the number of keystrokes required to write email on mobile devices such as smartphones. Existing devices can provide a frustrating user experience because they include small keyboards and context-insensitive typing corrections, which lead to errors and a poor email experience. We proposed to use our personal email as a source for both training and testing a model that predicts accurately the next typed word.

Our initial research suggested that we use a Markov model to compute $P(w_n\vert w_1\cdots w_{n-1})$ , where $w_n$ is the $n$ th word of an $n$ -gram model and that the Good-Turing method [2] may provide a suitable means for smoothing probability mass among seen and unseen $n$ -grams before predicting words.

In summary, we have completed the following tasks and appear to be slightly ahead of schedule.

• Collect, parse, and split email bodies into $n$ -grams, and compute statistics for use in the prediction model.
• Create the Markov-based, word prediction model.
• Analyze model performance and productivity increase in writing emails.

In the rest of this document, we will describe the status of our research. In Section 2 we discuss the dataset, its origin, characteristics, and preprocessing. Then, in Section 3 we discuss the probability mass estimator used to smooth $n$ -gram distributions and the technique used to predict words. In Section 4 we begin to analyze the performance of our approach, and in Section 5 we conclude by reviewing the schedule.