3 ms·

Perhaps this is just phase one, to build a dataset out of YouTube videos and sponsored promotions. Phase two could be training a classifier on the dataset. Then

• gpm 7y ago
  It would have to be quite the classifier, since the advertisers would be able to test against it and try to break it. False positives would heavily impact usability.
• crazygringo 7y ago
  On the one hand, classification might be a bit easier given that YouTube auto-generates transcripts -- so it only has to work on text. On the other hand, I just don't think it's possible -- it's going to have way too many false positives, because there are already so many unsponsored YouTube videos that review products or tell their customers to buy things, and there's very little actual text involved in the sponsorship segment. And you've got to 1) identify that there's a sponsored segment at all which is hard enough, but then 2) accurately detect where it starts and ends, which is even harder. And if this ever did manage to work on existing videos, it just turns into an arms race which neutralizes it -- sponsors will require content creators to be more creative in how they embed the sponsorship so that they only get sponsored as long as the segment makes it past the classifier.
• ajayyy 7y ago
  Sponsorblock developer here. The best part is that I made the database completely public, so someone else can do "phase 2" without me even having to be involved :)