What is the process of selecting relevant rows?

[HIEROT]

The paper proposes that we use N-gram overlap to select rows relevant to the statement/context to make a content snapshot. Is this done during prprocessing the dataset? Doesn't that mean this operation cannont be optimized?
I am a bit confused. Because feeding all rows of a chart into the model requests too much memory, I guess the rows are already selected before this.
It will be much appreciated if you can make this clearer.
Thanks!