3 ms·

You can use JSON-LD to cleanly embed RDF into LLM-friendly JSON. Not sure if there's a direct Markdown equivalent, but the Turtle plain-text syntax for RDF is v

You can use JSON-LD to cleanly embed RDF into LLM-friendly JSON. Not sure if there's a direct Markdown equivalent, but the Turtle plain-text syntax for RDF is very simple and modern LLMs should be able to cope quite fine with it.

alganet 1y ago
The purpose of RDF is to enable reasoning ("I'm sharing this with you, and I'm also sharing how to reason about this thing I shared with you, so we can all reason about it the same way"). If you show me an LLM that can take in serialized RDF and perform reasoning on it, I will be surprised. Once the LLM takes in the RDF serialization, it is a dead end for that knowledge: by principle, you can't rely on anything the LLM does with it. In a world of LLMs, it makes much more sense to put the semweb technologies alongside the training step instead. You create ontologies and generate text from triples, then feed the generated text as training for the model. This is good because you can tweak and transform the triples-to-text mechanism in all sorts of ways (you can tune the data while retaining its meaning). It doesn't make much sense to do it now, but if (or when?) training data becomes scarce, converting triples to text might be a viable approach for synthetic data, much more stable than having models themselves generate text.
- zozbot234 1y ago
  > If you show me an LLM that can take in serialized RDF and perform reasoning on it, I will be surprised It's easy, you just ask the LLM to convert your question to a SPARQL query, then you sanity-check it and run it on your dataset. The RDF input step is just so that the LLM knows what your schema looks like in the first place.
  - alganet 1y ago
    I don't understand how this can work. Can you make me a quick demonstration using a publicly available model and the dbpedia SPARQL endpoint? https://dbpedia.org/sparql https://dbpedia.org/sparql

alganet 1y ago

The purpose of RDF is to enable reasoning ("I'm sharing this with you, and I'm also sharing how to reason about this thing I shared with you, so we can all reason about it the same way"). If you show me an LLM that can take in serialized RDF and perform reasoning on it, I will be surprised. Once the LLM takes in the RDF serialization, it is a dead end for that knowledge: by principle, you can't rely on anything the LLM does with it. In a world of LLMs, it makes much more sense to put the semweb technologies alongside the training step instead. You create ontologies and generate text from triples, then feed the generated text as training for the model. This is good because you can tweak and transform the triples-to-text mechanism in all sorts of ways (you can tune the data while retaining its meaning). It doesn't make much sense to do it now, but if (or when?) training data becomes scarce, converting triples to text might be a viable approach for synthetic data, much more stable than having models themselves generate text.

zozbot234 1y ago
> If you show me an LLM that can take in serialized RDF and perform reasoning on it, I will be surprised It's easy, you just ask the LLM to convert your question to a SPARQL query, then you sanity-check it and run it on your dataset. The RDF input step is just so that the LLM knows what your schema looks like in the first place.
- alganet 1y ago
  I don't understand how this can work. Can you make me a quick demonstration using a publicly available model and the dbpedia SPARQL endpoint? https://dbpedia.org/sparql https://dbpedia.org/sparql