Back in the old days, it was common to write rule-based systems. Systems that do;

• data -> [rules] -> labels

Nowadays, it's much more fashionable to use machine learning instead. Something like;

• (labels, data) -> [model] -> rules

It might be a good time to ask ourselves, is this a better approach? Or have we lost something while in this transition? In this talk, I hope to defend rule-based systems and highlight why they, literally, rule.

The talk will consist of many examples around this topic. In particular:

• I'll do a demo with features from spaCy where I'll show how easy it is to come up with reasonable rules using their amazing Matcher object.
• I'll demonstrate how rules make it easy to iterate on data. I'll also explain why this is more important than iterating on models.
• I'll demonstrate a blind spot for deep learning models that rule-based approaches don't suffer.
• I'll explain how at Rasa, we've introduced that rule-based tools to our stack. These tools turn out to make life much easier when you're designing a chatbot.
• I'll explain how you can combine the ML way of thinking with rule-based systems and get a lot of safety in return.
• I'll conclude by announcing and demonstrating a new tool. It will be scikit-learn compatible and it might just make rule-based approaches a lot more powerful. It's like the Matcher from spaCy but for tabular data.

I hope that these examples make people rethink the way that they do their modeling. There's an amazing opportunity to add some natural intelligence back into our systems. This is great because this might just make them more accountable.

We need more systems that play by the rules.