via Aaron B.
This is really interesting. I’ll admit that some of the questions around accessing data have plagued our organization for sure as we have adopted our own KG, but my sense is that it remains a really powerful way to represent data that can’t be totally captured well in a relational model. I wonder if the bigger problem is less with the technology but rather that the nature of KG development has leant itself to more proprietary work that hasn’t made its way to open source communities
I dont know if I agreed with the assertion that Triples result in a lot of data modeling complexity that JSON/relational obscures. We have found that the RDF model has made it actually easier to add to the model nimbly without dealing with all the headaches that can come with relational models
Most of the points are highly questionable, subjective (a shame they provide no references) or ignorant e.g. ignoring the fact that JSON-LD is a first class serialization for RDF. But could someone help me out on understanding what "relation splurging" is? I think it's complaining about lots of unclearly-named object properties. But at least KGs allow you to define properties with a whole bunch of definitional schema and metadata, so you're not reliant on the name. And relate/reuse them (equivalentProperty, subProperty etc). That seems a lot better than relational database foreign key columns scattered across lots of tables with inconsistent or unclear names and no way of relating them. And don't get me started on field definition in (general) JSON!
WRT "splurging" I think the complaint was that triples promote a proliferation of different relations that all mean the same thing. And even more so when relations are combined from different sources.
However, with the growth of relations from different sources, the relations can be increasingly more ambiguous between each other, e.g., belongs_to, part_of, included_by, etc. As a result, it is just hard to disambiguate by neither human nor machine.
To that second point specifically - um ... yes? Sure, data triple-described data from different sources are going to have different relations defined that you need to normalize - but (to your point Pete) so is RDBMS data, or any other sort of data, and KG technologies actually have approaches that facilitate the normalized use of heterogeneously-acquired data.
Yes I have the same understanding with that "relation splurging". I think this comes from the NLP background of the autors. Methods for automated KG extraction from text often spit many relations made from the terms found in the text. There is as such proper, cleanly architected ontology at that point but rather a terminology of relationships. I wonder how NLP folks in this group are dealing with this.
I feel the blog argument is more towards any structured data assimilation/curation itself. The set of problems/hurdles posed by the authors also exist when you are creating a schema (or a set of schemas) for structuring your data in RDBMS. Actually even more. EKG as a solution is a step towards that.
I concur with the views expressed here. Reading the paper suggests to me a frustration on the authors part that there isn’t a singular precise model of an imprecise world, and that the expressiveness of KGs and similar only adds to the ambiguity and messiness they are struggling with. I think this points to a failure to appreciate that, for any domain, variety absorbs variety; that is, to capture the domain at any one level, the model needs to be at least as expressive as that level — thus you either accept loss of fidelity or a rise in complexity. Feels to me like they are complaining about a 3D world as seen through a 2D perspective 😕
Better have this discussion in the comments right under that article 🙂
