Which paper about best way to calculate contextual embedding ?

[lfoppiano]

There is a picture which is quite interesting in https://jalammar.github.io/illustrated-bert/

You wrote: "Which vector works best as a contextualized embedding? I would think it depends on the task. The paper examines six choices (Compared to the fine-tuned model which achieved a score of 96.4):"

Which paper do you refer to?
Either I overlooked or the reference should be added.

Thank you in advance
Luca