Explore Proximity of Embeddings to their Classes

Explore if proximity/distance in embedding space can be used at all and solely for classification and further explore what sorts of pre-processing can improve results. These can be simple or very advanced: POS tagging, knowledge augmentation, ….

Steps:
1. Use a benchmark where labels are on sentence chunks/noun phrases available
2. Use KNN to determine classification metrics
3. Vary: (a) use compressed/reduced embeddings (eg UMAP) (b) filter by NLP tags