Using ULMFiT for other tasks besides text classification

Are there examples (notebooks) of using ULMFiT transfer learning for other NLU/NLP tasks besides of text classification?
Like may be question/answering or some other?

1 Like

I’m in the process of working on a few with my project Stack Roboflow

I launched v1 a couple of weeks ago which was basically default ULMFiT fine tuned on Stack Overflow data and with a beefy custom regex to reconstruct the output.

I launched v2 last night which customized the tokenization process for the task and let me ditch the domain-specific parts of the regex. (I made it so the tokenizer is whitespace, punctuation, and camelCase aware)

Next I’m working on generating answers. And I have a few other features I’m hoping to add like contextual tagging and probability of answer acceptance.

Here’s an example of the output:
image

1 Like