Did you know Google Search uses some sort of question fringe score? This is not part of the Google search data leak from a year ago, but rather from the Google exploit that he found and got paid for by Google as part of Google's bounty program.
It is under the name question_fringe_score.
Mark Williams-Cook uncovered this new one and posted about it on LinkedIn. He said, "More unrevealed info from last year's Google exploit: Google calculates a "question_fringe_score" when processing user searches. What does it mean?" He also shaed this screenshot of question_fringe_score value for a specific query:
If you are wondering how Mark found this, he did an excellent presentation on that late last year - so check that out.
Mark's theory, as he wrote, "I can't find any direct mention of this in Google patents or docs, however my guess would be it is likely a score estimating how far a query (especially a question) sits on the 'fringe' of Google’s known entity/knowledge space and how atypical or long‑tail it is."
Przemysław Charchan in the comments has a theory on what it might be scoring. He wrote:
The hypothesis that "question_fringe_score" only pertains to the long-tail nature of a query is incomplete. According to the documentation, "fringe" is a system within the Safety category, not just a measure of a query's statistical rarity
- Query Classification: In the ClassifierPornQueryMultiLabelClassifierOutput module, the fringe label appears alongside other safety classifiers such as porn, violence, offensive, spoof, and vulgar. This places "fringe" within the context of quality and potential harm assessment.
- Content Evaluation: The QualityFringeFringeQueryPriorPerDocData module contains the encodedDaftScore attribute (Document About Fringe Topic), which evaluates whether the document itself is about a "fringe" topic. The system, therefore, evaluates the content, not just the query.
- Link to Misinformation and YMYL: In the same module (QualityFringeFringeQueryPriorPerDocData), there are attributes that directly link the "fringe" system to predictions regarding misinformation (encodedChardXlqHoaxPrediction) and YMYL content (encodedChardXlqYmylPrediction)
Forum discussion at LinkedIn.
