Google has a printed patent named “Generative summaries for search outcomes” which is believed to be the patent behind the Search Generative Expertise launch we noticed earlier this 12 months. This patent was filed on March 20, 2023 and permitted on September 26, 2023 below the patent ID US11769017B1.
Juan Gonzalez Villa posted a thread on X breaking it down, which I’ll embed under so you may learn it.
The summary reads:
Not less than selectively using a big language mannequin (LLM) in producing a pure language (NL) based mostly abstract to be rendered in response to a question. In some implementations, in producing the NL based mostly abstract further content material is processed utilizing the LLM. The extra content material is along with question content material of the question itself and, in producing the NL based mostly abstract, could be processed utilizing the LLM and together with the question content material—and even unbiased of the question content material. Processing the extra content material can, for instance, mitigate occurrences of the NL based mostly abstract together with inaccuracies and/or can mitigate occurrences of the NL based mostly abstract being over-specified and/or under-specified.
Listed here are Juan’s posts:
Earlier than we begin: there are at the moment no different patents assigned to Google mentioning generative AI methods and search.
There is perhaps different patents on the identical subject nonetheless pending, however I consider this doc is extremely related to Google SGE as it really works proper now. pic.twitter.com/L4Ly5Ows2c
— Juan González Villa (@seostratega) November 29, 2023
1. Receiving a question related to a shopper system. The question could be explicitly entered by a person or robotically generated based mostly on context.
2. Deciding on a set of search consequence paperwork (SRDs) which can be conscious of the question and associated or current queries.
— Juan González Villa (@seostratega) November 29, 2023
4. Processing the SRD content material snippets utilizing an LLM to generate LLM output. An elective summarization immediate can be included.
5. Producing a pure language abstract utilizing the LLM output. This leverages robustness of the LLM whereas constraining it to the SRD content material.
— Juan González Villa (@seostratega) November 29, 2023
This checks out with one thing acknowledged within the doc “An Overview of SGE”, made public by Google round SGE’s launch, though we did not have any extra particulars.
The patent now offers loads of element round how and why a number of fashions can be found and can be utilized: pic.twitter.com/U035N35Bxl
— Juan González Villa (@seostratega) November 29, 2023
3. The choice could be based mostly on:
• Processing the question with a classifier to foretell finest mannequin(s)
• Detecting sure phrases within the question to point appropriate mannequin(s)
• Contemplating search consequence varieties/content material to find out acceptable mannequin(s)— Juan González Villa (@seostratega) November 29, 2023
So by dynamically deciding on from a number of candidate generative fashions, the invention goals to optimize accuracy and effectivity by selecting essentially the most suited mannequin(s) for any given question.
The patent additionally present some particulars on how the hyperlinks to sources work:
— Juan González Villa (@seostratega) November 29, 2023
1. “A portion, of a visually rendered NL abstract, that’s supported by a primary SRD could be selectable (and optionally underlined, highlighted, and/or in any other case annotated).
A number of the portion may end up in navigating to a hyperlink similar to the primary SRD.”
— Juan González Villa (@seostratega) November 29, 2023
3. The hyperlinks could be normal hyperlinks to the SRDs or particular anchor hyperlinks to the parts that present the verification.
4. The linkified parts could be decided based mostly on evaluating the abstract content material to SRD content material utilizing encoder fashions to establish verified parts. pic.twitter.com/zlJFFO2c6Q
— Juan González Villa (@seostratega) November 29, 2023
This is how confidence works, in response to the patent:
1. Confidence measures could be generated for parts of the abstract or for the abstract as an entire.
The arrogance measures are then used to find out which confidence annotation from a set of candidates must be utilized.
— Juan González Villa (@seostratega) November 29, 2023
3. A textual “excessive confidence”, “medium confidence”, or “low confidence” annotation could be annotated for the NL based mostly abstract as an entire.
Every of the parts of the NL based mostly abstract could be annotated with a corresponding colour that displays a level of confidence in that portion
— Juan González Villa (@seostratega) November 29, 2023
Thanks for studying to date.
This is the hyperlink to the patent: https://t.co/AtGMiWqzG9
Now, I will clarify how I discovered about this patent, and one other fascinating factor:
What Google SGE and Featured Snippets have in widespread… 🤓⤵
— Juan González Villa (@seostratega) November 29, 2023
In response to his personal profession abstract, he labored in SGE at Google between Nov 2022 and Sep 2023. He’s now a Google Fellow, the best rung on the ladder for Google engineers, which they attain because of “constantly excellent accomplishments”. pic.twitter.com/aZmpfkajmI
— Juan González Villa (@seostratega) November 29, 2023
Their careers afterwards have been very related: they each went to Apple, the place they labored in Search and Siri, and got here again to Google on the finish of 2022 to work in SGE.
So… how did I discover the patent? I made a decision to search for patents by any of those two engineers and there it was. 💡 pic.twitter.com/1uUPi9myoj
— Juan González Villa (@seostratega) November 29, 2023
Their careers afterwards have been very related: they each went to Apple, the place they labored in Search and Siri, and got here again to Google on the finish of 2022 to work in SGE.
So… how did I discover the patent? I made a decision to search for patents by any of those two engineers and there it was. 💡 pic.twitter.com/1uUPi9myoj
— Juan González Villa (@seostratega) November 29, 2023
If you happen to loved this thread, any suggestions, likes and/or RTs are appreciated.
Additionally, this is a publish model of this thread:https://t.co/JSt9qNcVTr
— Juan González Villa (@seostratega) November 29, 2023
Good write up Juan!
Discussion board dialogue at X.