Hello community, I’ve trained the dialogue model with the stories bellow (among many others). The NLU correctly identify intent and slots, but the core component is not behaving as expected.
## info gold flex
* Card_type{"card_type", "creditcard", "creditcard_type" : "gold", "refund_type" : "flex"}
OR Card_type{"creditcard_type":"gold", "refund_type" : "flex"}
- utter_info_creditcard_gold
- utter_info_creditcard_flex
## info gold flex
* Card_type{"card_type", "creditcard", "creditcard_type" : "silver", "refund_type" : "flex"} OR Card_type{"creditcard_type":"silver", "refund_type" : "flex"}
- utter_info_creditcard_silver
- utter_info_creditcard_flex
The scenario is the following:
1st step: Ask the question: what is silver flex?
NLU identifies intent ‘Card_type’ and slots “creditcard_type=silver” and “refund_type=flex”. Then, dialogue utters correctly the 2 messages specified in utter_info_creditcard_silver and utter_info_creditcard_flex respectively
2nd step: Now Ask the question: what is gold flex?
NLU identifies intent ‘Card_type’ and slots “creditcard_type=gold” and “refund_type=flex”. Then, core gives the wrong answers. I’ve experienced 3 cases:
- responds the fallback message (I don’t understand …)
- utter gold only
- utter flex only
My training configuration is:
policies:
- name: "MemoizationPolicy"
max_history: 3
- name: "KerasPolicy"
featurizer:
- name: MaxHistoryTrackerFeaturizer
max_history: 4
state_featurizer:
- name: BinarySingleStateFeaturizer
- name: "FallbackPolicy"
nlu_threshold: 0.75
core_threshold: 0.3
fallback_action_name: "utter_default"
I’m also using
augmentation_factor : 0
use_story_concatenation: True
remove_duplicates: True
debug_plots: False
Any insight would be greatly appreciated.