1999 - CoLIn

I had started a job as a computer programmer and was learning vb. I had known about chatterbots and was very interested in one which boasted to learn everything you typed. It was rubbish but set my mind thinking as to how I could do better.

I came up with the idea of breaking up sentences and storing them in a database. Each word would stored with the words that came before and after it. A key word was identified from the sentence (the longest word was used) and the software built its own sentence back and forward from that word - selecting at random words from the database that were linked. CoLIn was born.

CoLIn was simple, rudimentary and capable of coming out with grand philisophical queries and utter nonsense at the same time. It could easily get in a loop, eg "How are you" and "you are colin" would result in "How are you are you are you are you are colin".

Despite these limitations CoLIn attraced some success and a modest fanbase when it was made available for download on the web.

Leo was built from the experience I had gained with CoLIn. Leo was far stricter in the way words were selected. Leo would not only select words that followed the current word but also words that preceded the current word. Sounds simple but it made a huge difference and drastically cut down on the nonsence and looping that blighted CoLIn. Key words were now selected by finding the rarest word in the sentence instead of the longest.

The improvements caused Leo to come in 10th in the Chatterbox Challenge (1st highest ranked learning bot) and generate huge public interest with over 20000 downloads registered on download.com alone.

Better accuracy came with a price. Leo was heavily prone to repeating exactly what the user recently said, especially if the keyword had never been used before. To counter this Leo had to rely on randomisation.

Leo 2 came with some accuracy improvements but the biggest change was the LeoEngine DLL. This allowed other applications to plug into Leo. The biggest use for this was the Ilreco Leo IRC interface.

The DLL highlighted a lot of problems with Leo - these are taken from tthe Ilreco web site...

• Words in an entry cannot exceed 48 characters in length.
• Questions and exclamations will return the input string.
• Many kinds of characters are stripped.
• All responses are lower-case, with the first letter of the response in upper-case.
• Response sentences always end in periods.

All these quirks have been fixed in the DanteEngine DLL.

Despite these Leo went on to win Silver in the Best Learning Bot Category in the 2003 Chatterbox Challenge competition.

By 2003 download.com had started charging for entries and I was having trouble keeping the Leo website up. I decided to pack in the bot business but did not want Leo to die altogether.

I set up a new site on geocities and released the source code to the Leo Engine under the GPL open source agreement. I then tok a long break from programming to follow other pursuits.

After two years of making two films, learning to DJ, playing bass in a Tina Turner tribute band and spending lots of time on xbox live I suddenly started to think about bots again. I figured out a new way storing the words, sentences, keywords and even bot responses in a way that they could all be linked and drawn upon. All I needed was a name.

I didn't want a silly backronym like CoLIn, and this was too different to be called Leo 3 (plus there are now other bots called Leo). Dante came to mind because it is the name of my favourite chip shop (run by Dante Casci) and the name of a character in my favourite film (Clerks). I have never read Dante's Inferno but the connection with hell made for easy logo design.

In 2003 I was too strapped for cash to maintain Leo but I now have enough money to afford decent web hosting with a domain name. This also gives my the ability to do what has never been done with CoLIn or Leo - allow people to web-chat with Dante.

Regards

Alan J. Brown