The Future of
Knowledge Graphs
in a World of
Large Language Models
Denny Vrandečić
New York, NY - May 11, 2023
“When in doubt, you can’t be wrong”
Disclaimer 1
CC-BY-SA 3.0, Pnautilus
CC-BY-SA 3.0, Pnautilus
Nothing in this talk is generated
*unless explicitly marked, or in a screenshot from an LLM
Disclaimer 2
This is a focused talk
Disclaimer 3
CC-BY-SA 4.0, Andrew Lih
CC-BY-SA 4.0, BrunelloN
~5 seconds
~5 seconds
0.56 seconds
~5 seconds
0.56 seconds
0.26 seconds
~5 seconds
0.56 seconds
0.26 seconds
~5 seconds
0.56 seconds
0.26 seconds
~5 seconds
0.56 seconds
0.26 seconds
Large generative model
Large generative model
Knowledge graph lookup
“ChatGPT-like Google would be 10x more expensive per query”
You can machine learn Obama’s birthplace every time you need it, but it costs a lot and you’re never sure it is correct
32 x 412 = � 15,129
Don’t get bedazzled by LLM’s capabilities, but use it where efficient
CC-BY-SA 4.0, EdoardoRamalli
CC-BY-SA 4.0, EdoardoRamalli
Stable diffusion
Public Domain, All images generated by Stable Diffusion
890,000,000
175,000,000,000
890,000,000
175,000,000,000
7,000,000,000
In a world of infinite content, knowledge becomes valuable
Overfit for truth
CC-BY-SA 4.0, EdoardoRamalli
author
it’s complicated
Superlative� Subject: Jupiter� Quality: large� Class: Planet� Locality: Solar System�
“$Subject is the $Superlative( $quality) $Class in $Locality.”�
Jupiter is the largest planet in the Solar system.
LLMs are awesome, but…
The future of Knowledge Graphs is brighter than ever
Thanks to a world with Language Models
Contact me
denny@wikimedia.org
That’s 9:34 am in New York time
Should LLMs store knowledge end to end?
A.I. TURNS THIS DATASET INTO A TEXT ON WHICH I CAN TRAIN THE LLM ON
A.I. EXTRACTS TRIPLES OUT OF THIS TEXT, SO I CAN QUERY IT EFFICIENTLY