In my case, I have numerical symbols for each intention. e.g:1 for affirm and 2 for deny. Or 3804 for Intention A and 3805 for intention B.
Per my testing, in “supervised_embeddings” pipeline, it can’t tell between numbers. I have training data 1 for affirm, and training data 2 for deny. But it will always pick intent affirm for me no matter I input 1 or 2. It’s likely to parse all numbers as same intention.
I know it this can be solved by non-NLP programming handling. But it will be better if I can use training data to achieve this.
I think this will not be possible via the supervised_embeddings pipeline. If your NLU data contain number and normal text for one intent, it will be impossible for the classifier to learn, that a certain number correspond to a certain intent. Numbers are also replaced by a unified token in the pipeline, so that the actual number is not seen by the classifier at all.