Skewed results when using ULMfit in a multi label problem

I’m using ULMfit in a multi label problem and having issues with the results I’m getting. I have a data set of around 1500 paragraphs of text, about 500 of which I’ve already labeled myself, which I’m trying to predict sentiment along with if the paragraph is an ‘idea’ or a ‘complaint’. I’m having problems with my predictions being almost 100% ‘true’ for the paragraphs of text being ideas and complaints, even though many of them definitely aren’t.

If anyone has any ideas as to why this may be happening and possible fixes to this issue I’d greatly appreciate it.


can you share your notebook here?

It was work on a project for the last company I was working for, so I won’t be able to share it.