- From: Graham Klyne <gk@ninebynine.org>
- Date: Wed, 14 Jun 2023 18:04:56 +0100
- To: semantic-web@w3.org
- Message-ID: <6225ec7d-b8fa-b459-e32c-94c6ac599fc9@ninebynine.org>
A friend was recently doing some experiments asking ChatGPT to generate some code, with the kind of mixed results you might expect. I suggested a strategy of asking it first to generate test cases, then asking it to generate code. This seemed to work, though some of the test cases offered were blatantly wrong, in a way that was obvious to a human reader. I'm wondering if this kind of strategy might apply to ontology- and data- generation? #g On 11/06/2023 21:40, Dan Brickley wrote: > On Thu, 9 Feb 2023 at 11:44, Dave Reynolds <dave.e.reynolds@gmail.com> wrote: > > There's already been some discussion here on ChatGPT and the extent to > which it can, or can't, do things like generate sparql queries and the > like; and people may be getting bored of the ChatGPT hype. However, in > case of interest, here's some notes on some lightweight playing with it > as an aid in writing simple ontologies: > > https://www.epimorphics.com/writing-ontologies-with-chatgpt/ > > tl;dr You can generate simple, superficial examples with it but it's of > limited use for practical work atm, though tantalising close to being > useful. Certainly don't trust it to do any inference for you > (unsurprising). OTOH getting it to critique a trivial ontology (that it > generated) for coverage of a domain was much better - so as an aid to > generating checklists of features of a domain to consider during > modelling it _might_ be of more use, even as it stands. > > > Thanks for sharing this. I know there is a tendency for people aligned with > Semantic Web to reject these technologies but in my view they bear close > scrutiny and are worth very serious attention. This doesn't mean we must like > everything about them, or they're the one road to [whatever].As a phenomena > this is an extraordinary turning point. > > This makes the ontology-authoring experiment quite interesting, since the > ground is shifting under our feet. As a community we have longstanding > debates, instincts, styles and differences on the question of how much to pull > into an explicit model, versus leave in scruffy text-valued fields (Dublin > Core vs FRBR, for example). So alongside using these new tools to help us > continue what we were doing before, they also raise questions about whether > new modeling habits will arise. The LLMs are better than anything prior at > unpacking the intent behind human prose - but at what point do we find they're > good enough to actually affect how we model things? Can we make ontologies > simpler and easier to use, without letting bias and weirdnesses creep in? > > Has anyone been experimenting with fine tuning in this context? SHACL/ShEx? > > > The step in the dialogues that really stands out, though, is when we asked > it to critique its own ontology. Its summary of features of organisations that > you might want to think about modelling was excellent. > > Very much agree on this point. Also wondering whether it could be useful as a > technology to make the more formal aspects of SW/RDF technology accessible to > non specialists (e.g. proofs, complex rules)... > > cheers, > > Dan > > > Dave > > -- Graham Klyne mailto:gk@ninebynine.org http://www.ninebynine.org Mastodon: @gklyne@indieweb.social GitHub/Skype: @gklyne
Received on Wednesday, 14 June 2023 17:05:06 UTC