Summarization for downstream task


I was wondering id anyone could point me to any work about summarization for a downstream task.

For example, given an NLP pipeline, one might want to first summarize the input and then perform some tasks (eg keywords extraction, classification, etc).
For very long input, a first summarization step makes the text more treatable. I know of groups / companies that do proceed in this way, in some cases.
However, one might want to directly summarize the text with the downstream task in mind: for keyword extraction, this might mean to keep as many keywords as possible, for classification to keep interesting features etc.

Is anyone aware of any research work in this direction? I have looked a bit and I did not find anything, but I would be surprised no previous work exists, so I am problably searching using the wrong keywords.

Any idea in this direction would also be highly appreciated :slight_smile: