Paper <<UserAdapter: Few-Shot User Learning in Sentiment Analysis>> proposed a method to create specific vector by fine-tuning bert in few-shot learning. In this article, we will introduce it.
Look at prcess below:
In this paper, it will generate some user-specific vector.
How to generate user-specific vector in few-shot learning?
Step 1: we will use pretrained bert.
Step 2: train user-aware model on massive user data
Step 3: fix bert and classifier parameters and only train user-specific vector.
From this paper, we can find:
(1) In order to make optimization stable, parametrization strategy is used.
Understand Parametrization Strategy in Deep Learning
(2) Users in bottom is also exist in middle.