I was reading this article:
And was wondering which response generation mechanism RASA uses as it seems it uses a hybrid between the two?
Thanks in advance.
The standard RASA is a pure ‘Retrieval-Based Model’. All bot utterances are saved in the domain (called templates) there.
However, when I looked into the code last time I found some hints to NLG stuff. Is this being planned or already in some kind of test phase?
Yes I agree but does it use pattern-based heuristics or intent classification using machine learning for the response generation mechanism?
That would be very interesting if they did introduce some NLG capabilities.
Depends on the classifier, you use. But other than the keyword classifier, they use machine learning.
Reponse generation is however a bit more complex. There are several competing response generators, called policies, which use different strategies. The MappingPolicy for example is strictly deterministic while others like the KerasPolicy use machine learning to determine the response.
1 Like
I see so if I’m using the Embedding Intent Classifier and the Embedding Policy then it is all machine learning based?