On Geodesic Distances and Contextual Embedding Compression for Text Classification

1 code implementation • NAACL (TextGraphs) 2021 • Rishi Jha, Kai Mihata

On one dataset, we find that despite a 12-fold dimensionality reduction, the compressed embeddings performed within 0. 1% of the original BERT embeddings on a downstream classification task.

Dimensionality Reduction General Classification +2