About Us

"Real stupidity beats artificial intelligence every time." — Terry Pratchett, Hogfather.

Convo is an application of supervised machine learning, with the aim of building a conversational system based on the principles of natural teaching and learning. In other words, we are aiming to develop a system that learns in a similar way to how humans learn a language: by listening to people and making its own attempts to copy them and express itself in that language.

The Convo project is a non-commercial project under the supervision of Dr David Hamill.

Here is a timeline showing how the Convo project has developed so far.

Chatterbots

Everyone knows about "talking computers", thanks to science fiction movies. However, the reality is that conversational systems are at a very early stage of development compared with fictional characters like the computer HAL in 2001: A Space Odyssey or the robot C3PO in the Star Wars series.

Conversational systems today try to emulate the complexities of human conversation, but they often fail badly in practice. This could be because they try to mimic superficial aspects of conversation but are not based on firm foundations. Despite initial appearances, they have no understanding of the conversations they engage in, so conversations can easily go wrong.

To find out more, read how most chatterbots work.

New Approach

Instead of starting with full adult human conversation, we're trying to develop a gradual learning scheme, via a series of experiments. In order to manage people's expectations of the software, the conversational system is represented by a character that people won't expect to be very intelligent or to speak good English: Ditto the donkey. Rather than trying to emulate the wide knowledge of an adult human, Ditto will only be able to converse on a limited range of topics. Hopefully this will expand as he learns more, under our supervision.

Ditto also has a limited emotional capability. Initially this will be one-dimensional, like a newborn baby, on a scale from happy to unhappy.

All the learning will be supervised and offline: we'll select examples for the system to learn from, train Ditto offline, then update the online version. Ditto won't learn directly from his conversations, although we'll use them as the source of examples. If everything goes according to plan, there will be a gradual improvement in Ditto's performance over time.

Contact

If you'd like to contact us, please email spam@spam.co.uk, replacing both occurrences of "spam" with "convo". ;-)