Over the course of one week, our chatbot has made fairly significant progress. It’s actually quite remarkable to think that Wilbur started with just one sample statement, and has gotten to a point where he’s identifying (in some cases) 100% of the health complaints thrown at him. In many cases the complaints are completely unrelated to anything he had seen before.
There was some concern initially, that the dataset generated by the class wouldn’t be large enough to help Wilbur make discernible progress, but it seems to have been unfounded.
Part of the reason for this success with such a small amount of training, is likely that Wilbur was given a very simple task to do, and no real sophistication in his understanding was expected. For example, statements like “I’m sleepy” were accepted as medical concerns, but that’s definitely context-dependent. Thus, Wilbur has a long way to go before he can be of any real help with health concerns, but as an illustration of AI learning, Wilbur has been a success.