Training with existing data

Does Rasa support any type of training / story generation from existing conversation data?

We have a bank of thousands of conversations from an old chatbot, and would like to use that to train our new Rasa bot. I saw a few ways to import data, assuming it was in a typical TrackerStore, but this data is not. We’ve just got a bunch of conversations in a MySQL database (but could easily convert it to JSON or some other format if necessary).

Thank you! Adam

Referencing the docs here: Training Data Importers

Could someone please help explain exactly what this does, and if it would help in my situation? It seems like I could write a custom importer to parse our existing conversation data and pass it in to Rasa… is this true?

Alternatively, would it be bad practice to use Rasa NLU to grab the intents for existing user messages and construct stories based on these?

It seems like I could write a custom importer to parse our existing conversation data and pass it in to Rasa… is this true?

That’s exactly what it does :slight_smile: basically it lets you customize the code that does the markdown(1.x)/yaml(2.x) parsing in order to parse your data instead.

Hi Adam! We’ve built a solution that does just that (and its free)! And integrates well with Rasa! HumanFirst allows you to upload tons of unlabeled conversational data in the form of 2-way conversations or utterances. You can use our ML-powered workflows to quickly explore your data and start creating and training intents. Once you’re happy with what you’ve created you can sync your work back to Rasa using our CLI. Check this article for more details.