The below article mentions "Some sentiments, such as greetings, escalation, thank you and help, are detected using artificial intelligence (AI) models and not using keywords. The AI models are more flexible and can learn based on the past usage"
I know we can use NLP/Text Analyzer for a chatbot and train the Text Analyzer but how to configure a self-learning AI Model in a webchatbot. The one which can learn based on past model.
Good question Piyush. NLP models in Pega need labels for classification (supervised learning). We need these labels to configure responses of chatbot. Unsupervised learning does not work with labels hence would not lend itself well to configuration where we need to control the response of bot.
An alternative to self-learning NLP model would be to use training data and automatic model builds. When some conversation in chatbot results in 'No match' or 'Multiple match', there is a corresponding feedback item created which is shown in 'Training data' tab of channels. You can chose to automatically build the model at regular intervals using this data.