I am wondering if there are more explanations about the vectors concatenation. In the paper, u, v, uv, and |u-v| are concat and then it is used for the Softmax loss. |u-v| seems to be intuitive since when u and v are close, this goes to near zero (semantically similar). But the meaning of u, v and uv are not clear, would you elaborate more? Also, I see some work took square of max(u, v), and it seems to be working well. Overall, I am curious why those are working and how they are designed.
I am wondering if there are more explanations about the vectors concatenation. In the paper, u, v, uv, and |u-v| are concat and then it is used for the Softmax loss. |u-v| seems to be intuitive since when u and v are close, this goes to near zero (semantically similar). But the meaning of u, v and uv are not clear, would you elaborate more? Also, I see some work took square of max(u, v), and it seems to be working well. Overall, I am curious why those are working and how they are designed.