Training Series: Create a Knowledge Graph: A Simple ML Approach

This talk will start with unstructured text and end with a knowledge graph in Neo4j using standard Python packages for Natural Language Processing. From there, we will explore what can be done with that knowledge graph using the tools available with the Graph Data Science Library.

Useful links:

[fredrikhansen]

This is so valuable. Thanks for creating the content, love it. Thumbs up!

[grahamconquer]

Audio lessons always clear loud asinorder pauses CLAP, KIS KIS Keep it simple keep it safe,

[nazarzaki]

Very helpful and interesting! Thank you

[mdazimulhaque]

Very useful. Thank you for creating this tutorial!!

[fernandopalacios]

It is an incredible and very clear explanation. Thanks.

[罗杰瑞-pg]

that's cool, thanks for your video

[DavidTatum-qi]

Would it possible to use GPT or other LLMs to scale the entity recognition across a lot of text data like emails and conversation transcripts? If not via GPT, how I go about extracting using information from raw text? Thanks

[ScriptureFirst]

40:00 really starts to get good. Would like chapters here forward

[henrymaguire]

I just wanted to point out, when evaluating your performance you say you have 76% accuracy which is better than 50% so automatically better than random (50/50 coin-flip). However, since your data is imbalanced with about 70% negatives, this is not a good baseline. It's only 6% better than a naive "everything is negative" baseline. It's still better though, so at least it is learning something!

[Alkis]

This is very well explained, but after playing around with notebook 02 I went and tried to create my own graphs using The first chapter of lord of the rings. It is only double the size of Obama's wiki page, but the code is extremely slow. It has been 1h45m and it still didn't finish making the svo vectors.
I heard about RAPIDS, a NLP library from nvidia that uses GPU and will see if it can speed things up a bit. This seems to be the perfect task for parallel computing.

How long did it take to make the obama example?
EDIT: It took me 3h45m to vectorize a text with 20k words in a i3 9400 (3.6Ghz). Since google colab terminates de kernel after sometime, you might want to do that one your own machine. Also, even though it exported all the nodes to neo4j, the relations are getting timeout error. I will try and export it as csv, like the one used in notebook 02.