Use word vectors to compute similarity of the page content to banned topics, and if it is > threshold, then avoid that page. When word vectors are trained on billions of words they can pick up even the slightest hint of pedophilia. They also understand slang and memes, if present in the training corpus. I think SOTA in content filtering and classification is pretty good today for both image and text.

Another way to filter pages is to follow it in time, for a longer period and if it has multiple positives, then ban it entirely.