One method I employed a lot "back in the day" for extractive summarisation of text, is TextRank [1], which applies that multi-billion dollar linear algebra procedure against text. I remember working at a news media company and testing TextRank against BERT for 2-3 sentence article summaries, and TextRank was better.