We understand that the old way of creating a Chatbot would have been to create a substantial repository of questions and possible answers to create the illusion of a conversation that could have been held between the user and the computer. This is static, labour intensive and not a self-learning system.
More recently (as we hear almost all the time) – even this morning, BBC Radio 4 announced that over Christmas, the Today show will be hosted by “AI”. Quite annoying that is so much in the news, however, for our purposed and for the objective of demonstrating AI as part of the “AI as a Tool for Art” topic to be presented at the end of term, I gingerly attempt to build/borrow/steal a working AI chatbot.
Without drowning in detail, a layered Recurrent Neural net is constructed as follows:
- Download a dataset (this could be movie subtitle files or twitter feed for example)
- Create a model
- Train the model
- Test the model
Tensorflow – using the python language, developed by Google seemed the most productive approach and I installed it first on my Windows machine but soon realised it was a lot easier to use Linux, so I use a Virtual machine to install all the necessary Python Libraries and with Github.com I used the code from suriyadeepan/easy_seq2seq to start my experimentation.
The example I chose was Mr. Donald Trump’s Twitter feed. I am already learning how to use seq2seq and will report back soon on my findings.