Classifying text using labeled BERT feature vectors

nlp
1

I have feature vectors created by BERT-as-a-Service (uncased large) (which is a tensorflow-based code) for 1000 short text each of which has one of 1 to 10 labels. How can I predict the labels using my own 1000 points dataset and feature vectors? Do you have a similar feature-vector based code you can point me to? Does it matter if these feature vectors are created by tensorflow?

2

It does not matter if feature vectors are created using TF, at the end they are all float values. The difference only comes to importing them to your code.

I don’t have the code but the general approach is

  1. Create a vocab for your data
  2. Create a dict or something for the embeddings
  3. Use your favourite model, while using nn.Embedding as the first layer in your model