It doesn't, which is exactly the problem. Ontologies inevitably have mistakes. When your reasoning is based on these "strong" graph links, even small mistakes can cascade into absolute garbage. Plus manual taxonomic classification is super time consuming (ergo expensive). Additionally, that assumes that there is very little in the way of nebulosity, which means you don't even have a solid grasp of correct/incorrect. Then you have perspectives - there is no monopoly on truth.

It's just not a good model of the world. Soft features and belief-based links are a far better way to describe observations.

Basically, every edge needs a weight, ideally a log-likelihood ratio. 0 means "I have no idea whether this relation is true or false", positive indicates truthiness and negative means the edge is more likely to be false than true.

Really, the whole graph needs to be learnable. It doesn't really matter if NMR is a chromatographic method. Why do you care what kind of instrument it is? Then apply attributes based on behaviors ("it analyses chemicals", "it generates n-dim frequency-domain data")