Intelligent Email

Paper Title - Intelligent Email: Reply and Attachment Prediction

Authors: Mark Dredze, Tova Brooks, Josh Carroll, Joshua Magarick, John Blitzer, Fernando Pereira

Comment: Jill

Summary:

This article deals with different types of email predictions. There are reply predictions, attachment predictions, and other tasks that can be treated as binary classification problems. Each email was represented by a sparse vector and in order for learning to occur they used a logistic regression classifier. In order to evaluate the system they used F-measure.

Reply prediction

This helps decide which emails need a user to respond to and allows the user to mark their messages as "replied" and "needs reply".

Each user needs to have a profile that is created from their past email responses and their contact list. If a user is CC'd in an email they are more than likely not required to respond. If the email was directly sent to someone regardless of people being CC'd it is more than likely required of them to respond. Using TF-IDF scoring on the words of an email it can determine the terms frequencies in order to figure out which emails have questions in them. Tests were done on four different participants and their emails.

Since the precision is higher than the recall it is indicating that there were many messages that were difficult to determine if they needed a reply or not.

Attachment Prediction

Not having a document attached to an email that requires the attachment to also be sent can cause chaos amongst the business workplace. When production depends on a document being edited in order to continue and someone forgets to attach it this could cause problems. Some suggestions for this would be to have the email client suggest possible things to attach from previous messages received by the same person.

Each prediction system could be useful for a work setting, but future works will be done on more complex user models in order to solve the problems better.

Discussion:

I thought this article was interesting because sometimes emails can get overwhelming. I know that I've waited for days for a professor or TA to email me back answers to a few questions I had for them and sometimes I wasn't able to continue work on an assignment until they responded. I think something like this, after being developed better, could possibly benefit the users of a specific email client. It may get people to respond quicker or remember to send the appropriate attachment when emailing someone. I know I've forgotten to attach something that was the initial point of sending the email in the first place.