I am thinking on different use cases for nlu. On would be to classify documents for easier lookups. For labeled data this would be just a multi-cat text-classifier, right? But this would could only detect keywords that have already been in the training set. Is there a way to train a model to suggest new keywords from texts?
This type of task is called as “KeyPhrase Extraction”. There are both supervised and unsupervised types available for this task. Check out this https://bdewilde.github.io/blog/2014/09/23/intro-to-automatic-keyphrase-extraction/ if you want to learn more.