mlp in language encoder

[zwx8981]

https://github.com/lichengunc/MAttNet/blob/de6f41ce5dc541425e5bd8ab96949ca09fb3de94/lib/layers/lang_encoder.py#L57

Hi, I find a mlp in the language encoder, which is used to project the original word embeddings into new representations. While it is not mentioned in the paper, so I am just curious about why you do this in your actual implementation? Is it a necessary operation to make the model work? How would it affect the model's performance?

Looking forward to hearing from you. Thank you so much.