Preparing training examples Watson Assistant

0
votes

I use Watson Assistant to classify clients emails.

I read a good answer about the best practices for creating intents Best practices for creating intents/entities with IBM Conversation Service. It gave me some ideas how to improve the results.

However, I have a question about cleaning examples before adding them to Watson.

Must I delete all numbers, names, dates etc.? Must I delete the phrases like “Hello, Thank you, waiting for your response…..”?

For example, the initial email is:

Hello, I’ll move to another city and my adresse will change after the 21 April. My new adresse will be London, ABC Street 123, app. 34. Could you update it in my profil? Thanks you in advance. Sincerely, Mark

The “cleaned” email:

I’ll move to another city and my adresse will change. My new adresse will be

Must I delete all unnecessary information or it is better to add the initial phrase to Watson?

Do you know what algorithms Watson Assistant uses to make a classification (intents)?

Thanks in advance for any response.

1
machine-learningibm-cloudwatson-conversation
Watson Assistant is a conversational service. You seem to try use it for something different. Wouldn't the NLU service be a better fit? ibm.com/watson/services/natural-language-understanding – data_henrik
The intents of Watson Assistant can be used like a classification tool. I don't use entities or dialog. NLU doesn't let me to classify the mails, but intents yes. – antonina

1 Answers

1
votes

You should not train on PII information. It adds nothing of value to your model. It also creates a risk if there is a data related audit. For example GDPR.

Find the essence of the mails intent and train on that.

Your example is that they want to update their profile.

Be aware that WA is not designed to classify emails. So you may need to design a mechanism to score each sentence, or handle multiple intents in an email.

Another option is Watson NLC, which I have seen used as a mail classifier.

You should also try to reduce PII being stored in Watson Assistant, although it is PII compliant, and supports GDPR.