when doing a formaction inside custom action there is fuctions for validating and last one def submit function after the slot is filled
Could you start a new question on that? Feel free to ping me in it but that way this thread can stay on topic.
Okay…but how would I ping you???..and I think I find some solution for that…but I have another question…this one is long to post but if u said so I would post
You can write the @ symbol in front of someones name and then they’re notified.
Examples; @koaning, @faiza_conte
@koaning hey as you specify some on spacy and white space tokenization …can we use both???or what is their difference between them?
All of our supported components are described on our documentation pages and you should be able to find answers there.
ya i did read them …so what i mean is if am providing enough training data i actually dont need spacy??..in this case i only need whitespace tokenization right??..do those models made in rasa house or ??
The spaCy tokenizer comes as-is. SpaCy is pre-trained and just like the whitespace tokenizer won’t behave differently if you give them more or less data. The spaCy models are described here while the whitespace tokenizer that Rasa provides is built on top of scikit-learn.
okay thanks
Hello @Ghostvv,
So if I have only 2 intents, I have to increase the value of ranking_length parameter to 2? Because the confidence returned for the predicted intent is too high even with a random input ( even when I don’t use a char n-grams CountVectorsFeaturizer component). Could you please explain this parameter to me and the scale_loss one?
Thanks in advance!
@Yasmine With Rasa Open Source 2.3 we have introduced 2 new parameters model_confidence and constrain_similarities. You can read about them in the changelog and what settings to choose from the migration guide as well as what they do.
TLDR: Set model_confidence=cosine and constrain_similarities=True in the configuration of DIETClassifier.
Let us know if this helps with the problem of high confidence for random input.
Hey @dakshvar22,
Thank you for your reply.
I did the following steps:
- Installing Rasa 2.3.0
- Adding the parameters model_confidence=“cosine”, constrain_similarities=True and loss_type: “cross_entropy”.
The confidence of inputs is decreased but the confidence of random inputs still high.
For example, the nlu output of the message “hey” which exists in training data gave the following results:
And the nlu output of the random message “ghjghjkgh” gave the following results:
Thus, the chatbot won’t return the nlu_fallback intent.
I hope you might be able to give me some insights on this.
Thanks in advance!
Can you please post your complete config as well?
Thanks! Can you try setting model_confidence to inner and compare the confidence values? The returned confidence values will be as raw as possible because they’ll indicate the actual similarity between input utterance embedding and label embedding. If still they are comparable to each other, we should start looking into the data and featurizers.
Hello @dakshvar22 ,
Sorry for the late reply.
I’ve tested all the model_confidence values, but always getting the same problem, even with linear_norm in rasa 2.3.4.
Regarding the data and featurizer, I put my project files here : Fallback doesn't work with 2 retrieval intents FAQ and Chitchat
I hope you could help me. Thanks in advance!