Seeing as it's been 8 hours without response, I'll provide some input, although not the "answer". I am aware that I can be a bit of a nay-sayer on this topic, so bear in mind that a lot of people are more positive on using GenAI for this.
My belief is that traditional methods still haven't been fully utilised and using AI for classification is harder and less accurate than just designing the process using ML and Transforms. And that if you can classify it first (from a previous run), don't cloud the algorithm with that. Similar to performing fuzzy matches.
Classification algorithms can be super bespoke as the terminology can be very unique for an industry/process. Hence, most AI implementations I've seen with this involve RAG. If using ChatGPT etc, you want your prompts to restrict/define the research set. So, prompts like "give weighting to terms used in the HVAC industry" or having your own list of terms and a prompt like "Use this list as the starting point and don't use more than 100 new terms". I haven't tested either of those prompts, but hopefully you get the idea.
I personally still think ML methods are better for classification, but it mostly depends on what your expectation is. If you are expecting to get a finished result from the prompt, then that gets a lot harder, and you have to be prepared to have incorrect classifications. Using NLP, we would normally get a list of terms that it closely aligns with and a value. Then that info can be used with other standard formatting to pick the best term. So, a mix of traditional methods and more advanced methods (NLP), essentially using the NLP as another source for decision.
You could always use the ChatGPT connector as another source, similar to the NLP above. Your main thing to design is what your finished list of classifications looks like. What do you want to achieve. Do you want 3% new terms everytime, or do you want to classify against a list...
