Advances and expectations on artificial intelligence have captured the public imagination and led to overly high expectations on chatbots, especially given the recent improvements in the speech-to-text and text-to-speech capabilities of computers. We are expecting AI to never be wrong and when a bot is not able to understand our intent we get frustrated.
Context is key
Talking to humans, we can count on the fact that they will remember what has been said previously in the course of the current and past conversations. This ability is what makes a conversation feel “natural”. This is “understanding the context” of a conversation.
Unfortunately adding such a skill to a chatbot is truly difficult because it’s almost impossible to anticipate every path a conversation may take.
If you use a rule based model to handle all the critical paths of a conversation it is likely that you’ll miss many of them or you will try to build an unmanageable state machine. All of this ending up in the same result: a bot saying “Sorry. I don’t understand”.
Advances and expectations on artificial intelligence, this is where Machine Learning comes to rescue.
This is where Machine Learning comes to rescue. While in rule based models bot responses are painstakingly hard-coded by software engineers using if-then rules, with Machine Learning algorithms, we can record a vast number of real life dialogues and answer the question “when the dialog was in a certain context, what did the human say?”. We can then teach our bot to behave like the human did.
As strange as it can seem, most of the more important bot platforms available on the market are still using rule based models to handle Natural Language Understanding (NLU).
At Athics we have made our choice and have embraced Machine Learning from the first day as the right tool to build our platform and drive its future evolution. We are also looking into Reinforcement Learning techniques to complement our solutions. But this is a big enough subject for a dedicated series of posts.