The target value vocabulary
Shuailong opened this issue · 1 comments
Shuailong commented
It seems that the vocabulary of the slot values are not added to the vocabulary, E.g., dontcare. Only the system and user utterance tokens are added. How does the model generate the tokens which are not from the dialog history context?
jasonwu0731 commented
The ideal way will be reserving a empty distribution for unknown words, similar to the implementation here.
We simplify the setting by adding all the vocabulary, and yes we should also add all the possible slot values (can possibly get a better result). For the "dontcare" value you pointed out, in our case we mainly generate that value using slot gate.
Hope this help.