How to calculate positive and negative based on the feature embeddings output by the model and their labels?