Concept-Embeddings using Gold-Scene Representation

[alphacoder01]

Hi,
In the paper, it mentions that
"To validate the correctness of our model design, we feed the operation modules with gold scene representation. Our CCO achieves 89.61%"

My doubt is whether using gold-scene representations leads to better learned embeddings of the concepts?
It'd be useful if you could explain this.

Regards,