TIL - Sentence Transformers

[Note: I'm going to start posting some shorter vignettes and thoughts on the blog rather than relegating them solely to LinkedIn (it is MY content after all - or maybe Google's; I should really read the TOS).  Previously, I'd wait to post pieces of 1K+ words but you people and bots are so lovely, I thought it'd be worthwhile to share my thoughts more often.]

I came across the nifty concept of Sentence Transformers while perusing the last few items of my favorite technology goody bag.  As AI marketing fatigue sets in (especially now that DeepSeek has proved to be the Alpha and Omega of all things AI - or maybe not), the term gets thrown around more and more for marketing purposes.  ANYTHING that 2 years ago would've been classified as either Machine Learning or Data Science (two terms that were extremely buzzy in the last decade but are now aging like Norma Desmond) has now been assimilated into the AI marketing family.  Bonus points if people make the Generative AI association sight unseen.

The problem with lumping everything under AI is that interesting problems or gains made outside of GenAI are painted with the same broad brush, for better or worse.  Sentence Transformers seem to fall into that category.

So, with that lengthy, potentially unnecessary preamble, I introduce myself (and at least a few of you) to Sentence Transformers.

The main concept is this:

# The sentences to encode

sentences = [

    "The weather is lovely today.",

    "It's so sunny outside!",

    "He drove to the stadium.",

]

print(similarities)

# tensor([[1.0000, 0.6660, 0.1046],

#         [0.6660, 1.0000, 0.1411],

#         [0.1046, 0.1411, 1.0000]])

If you're not a software engineer (or even if you are), don't worry too much if the code above makes sense.

In a nutshell, Sentence Transformers look at other phrases they're grouped with and detect if the meaning is similar.  In the case above the 1st sentence is exactly similar to itself (hence the 1.000 correlation rating - think of it as a percentage if it's confusing).

The 2nd sentence is also about the weather, so there's a 0.6660 correlation with the 1st sentence.  The 3rd isn't weather-related, so there's only a 10% similarity.

The 2nd and 3rd lines from the tensor simply show the other sentences similarities to their respective neighbors. 

I don't have any immediate applications for this, but I thought the concept (and the fact that there's a simple library for utilizing it) was neat and refreshing when everything being pushed in this space lately has been solely about GenAI.

Until next time, my human and robot friends.