Hi, Rasa community. I am trying to make a multilanguage bot, and for getting the utmost performance I need to separate Rasa Core and NLU models. However, I can’t make Rasa Core completely language-independent, the reason is I can’t train Core model without responses. Is there a proper way to separate these models or this is some issue?
P.S. I am using Rasa 2.5.0
You will have two separate sets of NLU training data (one for each language) and a common core/dialogue training data.
You’ll be running rasa train twice, once for each language, and this will produce separate model .tar.gz files.
You would probably have two config.yml files with different pipelines for the two languages.
I’ve worked with a setup that does this separation in a single repo and then uses scripts to run the rasa validation, train but first it generates a separate data-build directories with the language specific setup (button common rules/stories). The nice thing about this approach is that you have a single repo but the negative is that Rasa X does not recognize this configuration.
Greg
Thanks, @stephens for your reply.
I agree that this approach is a straightforward way to make bot multilanguage. But I want to create NLU model for each language and common dialogue manager. In my opinion, it is an optimal way to create a multilingual bot, by reusing the core model. If I am mistaken, please share your thoughts with me.
Yes, the approach I described creates separate NLU models and a common dialogue manager.