How can I ensure that both DIET-identified entities and Duckling entities are included in either the “entities” recorded by each event, or are both included in the slots filled by a form? It seems that DIET overwrites Duckling entities and their metadata when I fill slots with a generic time entity that DIET has been trained to predict. I want to be able to leave entities for time (for example) to be left in my training data so that I can test against them. In addition, Duckling doesn’t pick up certain time spans that I want to make DIET learn to recognize.
Generally, you don’t want to label the entities that duckling picks up in your training data. However, as you said there are spans, that duckling doesn’t pick up. So in that case you might include those, but you should probably use another entity name for those to be easier distinguishable. You could even think of not using duckling at all, if you got lots of training data. Now test the NLU alone first, slot mapping is a whole other topic. But if your NLU pipeline don’t pick these entities up they’ll never end up on a slot.
I’m not 100% sure, what your problem with slots is, but you can now map these entities to slots.
My question has to do with the metadata that Duckling generates that gives specific semantics according to its rules. It does not propagate to the slots. It would be advantageous for metadata discovered by Duckling to be adjoined to other entities of overlapping spans.
There’s nothing preventing you from putting it in a slot.