Re:Ask HN: How to make good use of Google Alerts?
by PaulHoule on Friday February 19, 2021 @08:56AM
(#26193151)
They still have Google Alerts?
Technologically GA was trying to do something which is still beyond state-of-the-art.
Relevance based search systems assign every query-document pair a numeric score and then show you the top scores.
We know how to make relevance scores where score A > score B means it is a good bet that document A is more relevant than document B.
What we don't know how to do is calibrate relevance scores to a probability (e.g. "the probability is 90% that this document is relevant") or some other score ("reading this post has an estimated $3.72 in business value")
If you had that kind of score you could set a threshold for relevant vs not-relevant and control the quality and volume of results you get.
Actually we can calibrate conventional ranking scores and you run into two problems: (1) the highest score corresponds to a 70% relevance probability, (2) the score gives equivocal results for many relevant documents, so if you aim for 50% relevance you will get awful recall (the median relevant document might have a relevance probability of 20% or so).
Technologically GA was trying to do something which is still beyond state-of-the-art.
Relevance based search systems assign every query-document pair a numeric score and then show you the top scores.
We know how to make relevance scores where score A > score B means it is a good bet that document A is more relevant than document B.
What we don't know how to do is calibrate relevance scores to a probability (e.g. "the probability is 90% that this document is relevant") or some other score ("reading this post has an estimated $3.72 in business value")
If you had that kind of score you could set a threshold for relevant vs not-relevant and control the quality and volume of results you get.
Actually we can calibrate conventional ranking scores and you run into two problems: (1) the highest score corresponds to a 70% relevance probability, (2) the score gives equivocal results for many relevant documents, so if you aim for 50% relevance you will get awful recall (the median relevant document might have a relevance probability of 20% or so).