Understanding the Term Embedding in Word and Graph Contexts

·May 04, 2021 08:13 AM

does someone know where the term "embedding" (in word / graph embedding) comes from ? it's pretty confusing when you're new to the field : made me think of something that deals with sub graphs... why not call it word / graph vectorisation ? poke Vinay K Chaudhri Vinay Chaudhri

👍4

9 comments

· Sorted by Oldest

Luis Daniel Cazaurang Hernandez
·
Bonjour Véronique, new to the field myself as well, I found this quite interesting article about Graph Embedding specifically. They give a good explanation about it all. I hope it helps.
Véronique G.
·
bonjour Luis Daniel, I quickly looked through the article, it is a quite technical explanation about how to do embeddings, which unfortunately I currently have no time looking into in detail. My question was at the terminology level : why name this after a term that suggests there is "embedding" involved when (as far as I get it) it's more about vectorisation ?
Bojan Božić
·
The thing is that vectorisation would be imprecise as real-valued vectors that encode the meaning of a word or subgraph such that the words or nodes that are closer in the vector space are expected to be similar in meaning. So the term embedding signalises this proximity in the vector space, while simply vectorisation would mean the creation of the space itsef.
🙏1
Véronique G.
·
🤔 ... I can't figure how "embedding" can signalise "proximity in the vector space" ... this could be because I'm not familiar enough with vectors or maybe I'm missing one of "embedding" 's significations (English is not my mother tongue), as I understand it it's about "nesting" or "integrating". can't see how it can signalise proximity ... 🤷‍♀️ thks for the input though
Bojan Božić
·
It has to do with dimensionality reduction, to get your (possibly sparse) high dimension vector space down to an embedded lower dimension and denser space (hence the proximity). I’m with you when it comes to naming things intuitively, but if I had to decide between being factual or intuitive, I’d go for the former, because embedding is the name for the proximity space and vectorisation is a too general term.
👍1
Bojan Božić
·
Glad to have a discussion on embeddings though at the conference 🙂
😁1
Véronique G.
·
"vectorisation is a too general term" I think this is the key : for newbies, it would be a better graspable approximation, but for vectors geeks, it's too general
Bojan Božić
·
I wouldn’t call myself a “vector geek”, but the point I’m trying to make is: these terms are not interchangeable.
👍1
Véronique G.
·
when I say "vector geek" I actually mean someone who knows more about vectors than I do, and that is not a lot ... 😉
👍1

Luis Daniel Cazaurang Hernandez
·
Bonjour Véronique, new to the field myself as well, I found this quite interesting article about Graph Embedding specifically. They give a good explanation about it all. I hope it helps.
Véronique G.
·
bonjour Luis Daniel, I quickly looked through the article, it is a quite technical explanation about how to do embeddings, which unfortunately I currently have no time looking into in detail. My question was at the terminology level : why name this after a term that suggests there is "embedding" involved when (as far as I get it) it's more about vectorisation ?
Bojan Božić
·
The thing is that vectorisation would be imprecise as real-valued vectors that encode the meaning of a word or subgraph such that the words or nodes that are closer in the vector space are expected to be similar in meaning. So the term embedding signalises this proximity in the vector space, while simply vectorisation would mean the creation of the space itsef.
🙏1
Véronique G.
·
🤔 ... I can't figure how "embedding" can signalise "proximity in the vector space" ... this could be because I'm not familiar enough with vectors or maybe I'm missing one of "embedding" 's significations (English is not my mother tongue), as I understand it it's about "nesting" or "integrating". can't see how it can signalise proximity ... 🤷‍♀️ thks for the input though
Bojan Božić
·
It has to do with dimensionality reduction, to get your (possibly sparse) high dimension vector space down to an embedded lower dimension and denser space (hence the proximity). I’m with you when it comes to naming things intuitively, but if I had to decide between being factual or intuitive, I’d go for the former, because embedding is the name for the proximity space and vectorisation is a too general term.
👍1
Bojan Božić
·
Glad to have a discussion on embeddings though at the conference 🙂
😁1
Véronique G.
·
"vectorisation is a too general term" I think this is the key : for newbies, it would be a better graspable approximation, but for vectors geeks, it's too general
Bojan Božić
·
I wouldn’t call myself a “vector geek”, but the point I’m trying to make is: these terms are not interchangeable.
👍1
Véronique G.
·
when I say "vector geek" I actually mean someone who knows more about vectors than I do, and that is not a lot ... 😉
👍1