Pooling method for TSADE

[labbeth]

Hi @nreimers @kwang2049 ,
First of all, thanks for sharing your great work on sentence transformers!

Regarding TSDAE implementation, I understood that CLS pooling method was used as it gave the best results, or at least almost the same results as Mean pooling with the advantage of keeping position information. But I was wondering if you have any theorical insight to explain this empirical result, knowing that:

• Mean pooling was considered as a better method for previous SBERT implementations (right?)
• I don't really get the fact that position information is useful for this training

Thanks in advance!
Thomas