On the Wisdom of Opaque Identifiers

Wikidata uses opaque identifiers for its catalogued information resources. For example, the statement “wd:Q42 wdt:P69 wd:Q691283” may map to the label sequence “‘Douglas Adams’ ’educated at’ ‘St John’s College’” with a language-locale preference of English-US.

Opaque naming is wise for internationalization. However, in local, specialized contexts, e.g. the JSON-LD context for a particular dataset, there is value in human-readable suffixes so that e.g.

{
 "@context": {
   "d": "http://example.org/data/",
   "p": "http://examples.org/properties/"
 },
 "@id": "d:DouglasAdams",
 "p:educatedAt": "d:StJohnsCollege"
}

would be a valid JSON API response that is both human- and machine-actionable.

So, let’s continue to pursue “global” semantic names, right? This should be fine for scientific research, where we’ve more or less settled on use of the English language, no?

After reading Levary et al. (2012),1 I think that the pursuit of “global” semantic names for things is futile, particularly in the context of scientific research.

Words/phrases (aka “signs”) are often polysemous – they accumulate senses. For example, I inherited the infuration of my PhD advisor over use of the the term “resolution” in lithography/metrology to mean “minimum discernable feature dimension”, e.g. linewidth, versus what I consider to be the preferred sense of the term: “minimum discernable inter-feature separation”, e.g. line spacing.

When encountering terms of art in scientific and technical literature, an author can mean several things by it, even within a specialized discipline or sub-discipline. In fact, I think that use of polysemous terminology is to be expected in the research literature, as pursuit of new knowledge is a primary goal. For new knowledge to be accepted among practitioners, it must be acknowledged, and this may take the form of attempting a change or narrowing of the typical sense of of a term within the practitioners’ context. This attempt may be acknowledged by the community, and thus become (part of the pool of) knowledge it stewards, or it may be rejected.

Senses also accumulate signs. A concept (sense) may be expressed via a set of synonym words/phrases, aka a synset. Above, I tried to express a concept with the synset “{resolution, minimum discernable inter-feature separation, line spacing, …}”. The longstanding practice of authority control in library science, e.g. authority files and subject headings, acknowledges this accumulation, but nevertheless often persues a single, canonical display of each name, a canonical sign for each sense.

My takeaway from Levary et al. (2012)1 is that the appearance of a new concept in a dictionary often corresponds to the appearance of a word A “that was not definable at earlier times”, i.e. a word that required another new word B to define it, which in turn – being a new word itself – requires another new word C to define it, perhaps simply the former new word A, i.e. C is A.

In other words, new concepts are loops. It is not reasonable to assume that a concept corresponds to a single “node” in a word graph, a synset, and to choose a canonical sign from among candidate synonyms. Rather, a concept often comprises two or more (but not many more; loops with length > 5 tend to represent semantic misinterpretations) nodes.

This makes it even more difficult to stick a canonical semantic sign on a concept. Thus, I appreciate the wisdom of opaque identifiers, which can be mapped for particular (json-ld) contexts to identifiers with human-readable suffixes, with e.g. owl:equivalentProperty/owl:equivalentClass/owl:sameAs assertions, so that local human-actionable vocabularies can nonetheless be unambiguously unified.

Subscribe to get short notes like this on Machine-Centric Science delivered to your email.

References

  1. D. Levary, J.-P. Eckmann, E. Moses, and T. Tlusty, “Loops and Self-Reference in the Construction of Dictionaries,” Phys. Rev. X, vol. 2, no. 3, p. 031018, Sep. 2012, doi: 10/gffcxd↩︎ ↩︎