You might need something like salience ranking. And you don't need a heavy modern transformer model to do it. You can get by with yesterday's algorithms. This is LexRank. I think it's interesting because it uses graph-based centrality ranking, in a similar way that PageRank did.

Edit: Or create sentence embeddings, and cluster them or rank their distance against an input sentence, or something.

https://www.cs.memphis.edu/~vrus/teaching/nlp/papers/summari...