SUGGESTION: Module for RDF2VEC.

[bobvanluijt]

It might be interesting to see if there is a Weaviate module opportunity for RDF2VEC. Especially because the Weaviate schema is RDF-inspired.

Somewhat related: #1522

[bobvanluijt]

CC: @HeikoPaulheim

[etiennedi]

The major requirement for a vectorizer module is that it can independently assign a vector to each object (or a node). E.g. in the case of the transformers module if I have two sentences "My name is John" and "Jane likes ice cream", the vectorizer sees each object in isolation and can assign a vector to each.

If this is something that works with rdf2vec then it it's a great fit for a module. (I don't know - as I haven't looked into rdf2vec thoroughly yet). If, on the other hand, the entire "complete" graph is the input for the rdf2vec and the output is all vectors at once then it wouldn't be compatible with the "vectorize [individually] at import time" philosophy of Weaviate modules.

Having said that it would still work well with Weaviate natively, just not through a module. In this case the process would be:

1. Take the entire graph
2. Vectorize it outside of Weaviate, so that each node gets a vector
3. Import each node with the vector positions into Weaviate without a vectorizer module

In this case a good place to make the process easier for the user might be the python client?

[bobvanluijt]

Thanks for the writeup @etiennedi

If this is something that works with rdf2vec then it's a great fit for a module.

This would be the case, isn't it @HeikoPaulheim?

[HeikoPaulheim]

It's more of the second case. RDF2vec takes an entire RDF graph and outputs vectors for each node.

[bobvanluijt]

It's more of the second case. RDF2vec takes an entire RDF graph and outputs vectors for each node.

Interesting, in principle this should fit but maybe not as a module.

I'll come back to this, thanks Heiko!

[ali1k]

I also have a question related to RDF2vec:
in the 2nd case you described, if vectorization is done outside Weaviate for RDF data and then we use Weaviate vectorizer for some other type of data, would they still be compatible? or does it mean that RDF vecs cannot be used in conjunction with internal Weaviate vec?!

[HeikoPaulheim]

Not directly I guess. Embeddings created using different methods are inherently incompatible. However, there are quite a few approaches for learning mapping functions between embedding spaces. Am 05.05.2021 um 14:43 schrieb Ali Khalili:

I also have a question related to RDF2vec: in the 2nd case you described, if vectorization is done outside Weaviate for RDF data and then we use Weaviate vectorizer for some other type of data, would they still be compatible? or does it mean that RDF vecs cannot be used in conjunction with internal Weaviate vec?! — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1529 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA3XPPCM3TRW4PPS57WLFMTTME4QXANCNFSM42SWQRSQ>.

-- Prof. Dr. Heiko Paulheim Data and Web Science Group University of Mannheim Phone: +49 621 181 2652 B6, 26, Room B0.22 D-68159 Mannheim Mail: ***@***.*** Web: www.heikopaulheim.com

[bobvanluijt]

Well - in principle it should be possible. You could have Weaviate classes that represent the data and the RDF vectors and Weaviate classes that use -for example- a text vectorizer.

It's a bit abstract but a test setup should be doable.

[etiennedi]

To avoid confiusion, we need to distinguish two separate cases here:

Case 1: Single Weaviate - linked classes

This is the approach that @bobvanluijt describes. You'd have multiple Classes in Weaviate. Each class can have it's own vector space. They need not be compatible. Essentially you'd import your dataset twice, once with Text-Vectors, once with RDF-Vectors. You can then make graph-relations between the two classes. Note: These relations are not your regular edges that you already have in your dataset, these are additional cross-class edges which link the node in space 1 to the equivalent node in space 2. So whenever you do a vector search in one space, you can follow the link to the other vector space and perform a second search around the linked object. We also use this approach when linking images to text when we are using separate vectorizers for each.

Case 2: Single Weaviate - single Vector Space

This is the approach that @HeikoPaulheim describes. You have two separate vector spaces, then you train a mapping function to combine them into as single vector space (So far all of this has happened outside of Weaviate). Once you have a single model you can import it as a whole into Weaviate.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.