I claim that none of the commonly used embedding methods capture any semantics.
It's fine if you want to move from a symbolic to a numeric or geometric representation, but when you do, don't throw the semantic baby out with the symbolic bathwater.
I argue that a useful definition of semantics is "predictable inference". This makes it possible to have semantics outside a logical framework.
A methodological warning from 1976: don't fool yourself that wishful mnemonics in your knowledge graph are "semantics". Therefore, knowledge graphs without a schema/ontology is just a data graph, without much semantics.
Finally, a discussion of some embedding methods that do manage to take semantics into account (TransOWL, ball embeddings like ELEm and EmEL++, and box embeddings like BoxEL and Box^2EL.
So: even if you do move to a non-symbolic representation (numerical, geometric), make sure you keep the semantics: don't throw the semantic baby out with the symbolic bathwater.
Analytical Profile of Coleus Forskohlii | Forskolin .pptx
The K in "neuro-symbolic" stands for "knowledge"
1. The K in “neuro-symbolic”
stands for “knowledge” *
Frank van Harmelen,
Learning & Reasoning Group
Vrije Universiteit Amsterdam
Creative Commons License
CC BY 3.0:
Allowed to copy, redistribute
remix & transform
But must attribute
1
* With thanks to Wouter Beek
2. Highlights
• Bluffer’s guide to KG embedding
• What is semantics?
• What is not a knowledge graph
• A very fishy picture
• A lesson from 1976
• A baby in a bath
• Some hope for the K in neuro-symbolic
2
3. The K in “neuro-symbolic”
stands for “knowledge” *
Warning: no generative AI, no ChatGPT, no Large Language Models….
9. prediction
algorithm
9
Some things are better done geometrically (and not symbolically):
• Link prediction
• Node classification
• Relation extraction
Bluffer’s Guide to KG Embedding:
From Symbols to Vectors and back again
10. 10
Bluffer’s Guide to KG Embedding:
Different ways to compute the embeddings
TransE: |h+r-t|
RotateE:
symmetrie
<h,r,t> and <t,r,h>
composition
father’s mother =
mother’s father
11. Claim
None of the commonly used embeddings
capture any semantics
What is “semantics”?
11
15. Artificial Intelligence meets natural stupidity,
Drew McDermott, 1981
15
Wishful Mnemonics
A major source of confusion in AI programs is the use of
mnemnonics like “UNDERSTAND” or “GOAL”. If a
programmer calls the main loop of their program
“UNDERSTAND”, they may mislead a lot of people, most
prominently themselves.
What they should do instead is refer to this main loop as
“G0034” and see if they can convince themselves or
anyone else that G0034 implements some part of
understanding.
It is much harder to do this when using terms like “G0034”.
When you say UNDERSTAND(x), you can just feel the
16. Artificial Intelligence meets natural stupidity,
Drew McDermott, 1981
16
As a field, AI has always been on the border of
respectability,
and therefore on the border of crackpottery.
……..
In this paper, I have criticised AI researchers (including
myself) very harshly. To say anything good about anyone is
beyond the scope of this paper.
Prescription medicine for every AI researcher:
In order to maintain your mental hygiene, read
“Artificial Intelligence meets natural stupidity”
once yearly
20. So what is semantics for your computer?
Frank Bussum
birth-place
21. So what is semantics for your computer?
Frank Bussum
has-birth-place
domain: person
range: location
Frank is person
lowerbound
Has-birth-place
22. So what is semantics for your computer?
Frank Bussum
has-birth-place
domain: person
range: location
Frank is person
Has-birth-place relates
min-cardinality: 1
max-cardinality: 1
Bussum = Meren
lowerbound upperbound
Meren
Has-birth-place
26. Claim
26
None of the commonly used embeddings
capture any semantics
Because none of the commonly used embeddings respect
any of the reserved symbols from RDF Schema or OWL.
Embeddings do “distributional semantics”, but
predictable co-occurrence ≠ predictable inference
27. Claim
None of the commonly used embeddings
capture any semantics
Because none of the commonly used embeddings
can represent universal quantification
(and that’s where the inference comes from)
Embeddings do “variable free sentences” only,
and those don’t allow for any inference. 27
has-birth-place
domain: person
range: location
28. This is not a knowledge graph
28
It is a data graph
because it doesn’t support any inference
and therefore doesn’t have any semantics
30. Make embeddings semantic again!
(Outrageaous Ideas paper at ISWC 2018)
Abstract
The original Semantic Web vision foresees to describe
entities in a way that the meaning can be interpreted both
by machines and humans. [But] embeddings describe an
entity as a numerical vector, without any semantics
attached to the dimensions. Thus, embeddings are as far
from the original Semantic Web vision as can be. In this
paper, we make a claim for semantic embeddings.
Proposal 1: A Posteriori Learning of Interpretations.
Reconstruct a human-readable interpretation from the
vector space.
Proposal 2: Pattern-based Embeddings.
Use patterns in the knowledge graph to choose
human-interpretable dimensions in the vector space.
30
Neither of these are aimed at predictable inference
-> no semantics
31. From TransE to TransOWL
(and from TransR to TransROWL)
31
TransOWL:
TransE
Loss function
Summed over
all triples
32. More radical idea:
use more of the geometry
to capture the semantics
32
Male
Father
𝐹𝑎𝑡ℎ𝑒𝑟 ⊑ 𝑀𝑎𝑙𝑒
Parent
𝐹𝑎𝑡ℎ𝑒𝑟 ⊑ 𝑃𝑎𝑟𝑒𝑛𝑡
Spheres: ELEm, EmEL++
Male
Parent
Father
Boxes: BoxEL, Box2EL
𝑃𝑎𝑟𝑒𝑛𝑡 ⊓ 𝑀𝑎𝑙𝑒 ⊑ 𝐹𝑎𝑡ℎ𝑒𝑟 ?
34. Highlights
Bluffer’s guide to KG embedding
What is semantics?
What is not a knowledge graph
A very fishy picture
A lesson from 1976
• A baby in a bath
• Some hope for the K in neuro-symbolic
34
38. Symbolic loss function
during training
See survey of 100+ systems in Von Rueden et al, Learning, 2019
flower?
cushion?
“Parts of a chair are:
cushion and armrest”
“Given the context of chair,
a cushion is much more likely
than a flower”
P(cushion|chair) >> P(flower|chair)
41. Takeaways
41
Some things are best done symbolically,
some numerically, some geometrically
The Semantic Web community is now truly a
“learning and reasoning” community
Semantics ≠ symbolic representation.
Instead: semantics = predictable inference
Whichever representation you choose,
make sure not to loose the predictable inference
There is more to neuro-symbolic methods
than embeddings
Editor's Notes
DEZE klopt toch niet?
Zie eerder.
Mind-reading game to explain semantics.
If I show the audience the top triple, and we share a little bit of background knowledge in the square box (“ontology”), I can predict what the audience will infer from the top-triple. The shared background knowledge forces us to believe certain things (such that the right blobs must be locations) , and forbids us to believe certain things (such as that the two right blobs are different). By increasing the background knowledge the enforced conclusions (lowerbound on agreement) and the forbidden conlusions (upperbound on agreement) get closer and closer, and the remaining space for ambiguity and misunderstanding reduces. Not only misunderstanding between people, but also between machines.
Slogan: semantics is when I can predict what you will infer when I send you something.
Mind-reading game to explain semantics.
If I show the audience the top triple, and we share a little bit of background knowledge in the square box (“ontology”), I can predict what the audience will infer from the top-triple. The shared background knowledge forces us to believe certain things (such that the right blobs must be locations) , and forbids us to believe certain things (such as that the two right blobs are different). By increasing the background knowledge the enforced conclusions (lowerbound on agreement) and the forbidden conlusions (upperbound on agreement) get closer and closer, and the remaining space for ambiguity and misunderstanding reduces. Not only misunderstanding between people, but also between machines.
Slogan: semantics is when I can predict what you will infer when I send you something.
Mind-reading game to explain semantics.
If I show the audience the top triple, and we share a little bit of background knowledge in the square box (“ontology”), I can predict what the audience will infer from the top-triple. The shared background knowledge forces us to believe certain things (such that the right blobs must be locations) , and forbids us to believe certain things (such as that the two right blobs are different). By increasing the background knowledge the enforced conclusions (lowerbound on agreement) and the forbidden conlusions (upperbound on agreement) get closer and closer, and the remaining space for ambiguity and misunderstanding reduces. Not only misunderstanding between people, but also between machines.
Slogan: semantics is when I can predict what you will infer when I send you something.