3 ms·
I would indeed say that relevance measurement is the bias. You can build a bad search engine with simple rules like TFIDF, but pretty much anything more than th
by danpalmer 2y ago
I would indeed say that relevance measurement is the bias. You can build a bad search engine with simple rules like TFIDF, but pretty much anything more than that needs opinion and bias.
For example, if there's a content farm that rips off content from another site, assuming you don't want to show both the original and the farmed content (which would be a bad search engine by most people's metrics), you need an opinion on which the original was. Sometimes you might know for sure, but often you won't. Even if you do know who the original was, you need the opinion that the original is better than the copy. When you have content farms providing summarised versions, now you need an opinion on whether summarised content has value instead of/alongside the original – summaries do provide value, but how much?
None of this gets to politics or the legal requirements for handling things like hate speech or other content that is illegal in some/all regions, and you've already got essentially editorial control being exerted.
Social media moderation and search engine results are fundamentally the same problem, it's just that the base interaction is a search rather than viewing a following feed or viewing a video.
- janalsncm 2y agoYes, exactly. If a group of anarchists puts up spammy websites optimized for TF-IDF, blocking or downranking them suddenly means you are biased against anarchists. The flip side is that history is littered with examples of rules which seem neutral on their face but which have very biased intentions. For example “literacy tests” for voting, or poll taxes.