Levenshtein distance: obscure formula invented in 1965 for computing the difference between two strings of text and soon a part of the intelligence engine that drives Mailroom.
One of the more exciting promises of Mailroom has always been its ability to automatically suggest replies to your emails based on your past responses. Our first version of this did OK at first but would become less useful over time.
We spent some time researching why this was the case when designing Mailroom 2. One of the things we found was that people would generally use suggested replies as a starting point rather than sending the suggested email outright. If you made any changes, though, Mailroom thought you had started a new reply, so it never built up any good learning history.
For Mailroom 2 we did a couple of things to make this better. The first thing we did was we added the helpful new Suggested Reply picker. If Mailroom doesn’t automatically suggest the reply you want, this picker makes it easy for you to go find the one you do want. Doing this helps Mailroom get better in the future, so its always better to use the picker instead of typing a new email from scratch.
The second thing we did with Mailroom 2 was to add a bit of code that would look at the reply you sent and compare it to the suggested reply you chose. While we send whatever reply you type, if the reply you send is changed only a “little bit”, then Mailroom will not create a new saved reply; instead it will just train on the reply you chose, as if you had made no changes at all. It’s very clever.
Incidentally, the formula we currently use to determine when a reply has been changed a “little bit” originally came from the University of Pennsylvania. Was was invented to help some professors catch students who were turning in copied papers from past years. It turns out this formula works better with longer text (like term papers) than most emails, which brings us to the Levenshtein.
The Levenshtein algorithm is much more precise than what we use now. In fact, it is commonly used today to suggest alternate spellings when you fat finger words into search engines. It does even better on emails one or two paragraphs long, which is basically what most of our customers do, so we think its going to be much smarter.
This new formula is implemented and in our testing environment now. You should see it go live when we do our next roll out in a couple of weeks.
Plunging the depths of obscure mathematical algorithms; just one of the many things we do everyday to make your Mailroom experience just a little better. ;-)

subscribe to blog
listen to podcast
browse our archives
send us email




