Discovering Chunks in Neural Embeddings for Interpretability