Sounds unlikely, but that was the basis for a remarkable talk by Bob Kasenchak (co-authored by Ahren Lehnert) at an ISKO Meetup (14 June 2022). The talk began in uncompromising fashion with the cover of a famous work by French postmodernist philosopher Gilles Deleuze and psychoanalyst Felix Guattari, A Thousand Plateaus (1980), which is actually part two of a larger work, Capitalism and Schizophrenia. Perhaps mercifully for the audience, Kasenchak didn’t try to summarize the original, but restricted himself to just one analogy from the French work; but this analogy was powerful enough to justify the entire talk.
The French writers explore the difference between “rhizome” and “tree” thinking, or what you could call “rhizomatic thought” (bear with me). A tree grows hierarchically and vertically, typically from a single stem. A rhizome, in contrast, grows from a network of horizontal, lateral stems under the ground. Rhizomes don’t have a single stem, but everything is linked to everything else (nicely depicted in the presentation as “and … and … and”). This analogy works well for describing the difference between a standard taxonomy, which is hierarchical, and a knowledge graph, which is more lateral.
Extending the analogy, hierarchies tend to be copied by professional taxonomists in the form of an industry-standard, domain-specific taxonomy, which Kasenchak and Lehnert describe as “traced”: each version is a copy of the original. In contrast, knowledge graphs tend to be “maps”, each one original, and capable of growing in any direction. There is an excellent example of a rhizome-like map in the form of a subway map. Each station is linked to every other station, but there need not be any kind of single root. Lines can be in any direction.
Of course, analogies cannot be taken too far, but this simple idea proves remarkably powerful. Kasenchak’s description of the limitations of traditional taxonomies was all the more powerful coming from someone who spent several years working in a traditional taxonomy company.
One limitation of traditional taxonomies is about related terms. We all know that dogs are examples of pets, and dog food is an example of pet food. We also instinctively know that there is a relationship between “dog” and “dog food”: dog food is food designed for dogs. Similarly, a microscope is an example of a laboratory instrument, and a microorganism is an example of an organism. A microscope is used to see microorganisms. But the relation between “dog” and “dog food” is not the same as the relation between “microscope” and “microorganism”; unfortunately, in traditional taxonomy, all we have is the catch-all “related term”. In other words, hierarchical models are not good are distinguishing different kinds of relatedness.
The presentation then continued by showing one of those semantic diagrams, full of arrows, which we have seen so often in presentations about knowledge graphs. Their example was actually taken from a completely different presentation, but serves as a typical example. It is a map of a dog and its owner:
In this diagram, “Tom” is an instance of a dog. A dog is a mammal. A dog has fur. Tom, the dog, is owned by Rashan. Dogs like bones (and so on). You can see immediately it would be difficult if not impossible to capture all this information via a hierarchical taxonomy. I can understand that in a knowledge graph, the relationship is as important as the things connected.
What is not revealed, however, in this diagram, is the vast metadata requirement to specify the kind of relationship involved in each of these arrows. We all know that a major limitation of a traditional taxonomy is the sheer labour involved in building it in the first place, then maintaining it as new content is added. Yet, for the kind of knowledge graph above, equal or even greater labour is required if all the relationships are to be specified in such detail.
There was no mention in the talk of the effort involved in defining the triples, but perhaps as a silent answer to this objection, the next example was relatively simple – and didn’t require any triples. This was the JSTOR Text Analyzer tool, which has been available for some years on the JSTOR site. It comprises a blank box into which a user can paste any chunk of text or document URL. The system then identifies concepts from that text, and finds related content from the JSTOR corpus:
As an example, I uploaded the text from Wikipedia for the Gaia Hypothesis of James Lovelock. The system automatically identifies a number of “prioritized terms”.
If you wish, you can now amend the concepts chosen by the Analyzer, but of course, this requires some quite detailed subject knowledge, which many (perhaps most) JSTOR users will not have. Further down the screen, the system then identifies related content in JSTOR.
Unfortunately, the first related article found by the JSTOR Text Analyzer is completely irrelevant – this is an article about Classical Greek philology. Still, the idea is good, and most importantly, there is no need for the user (or anyone else) to identify the kinds of relatedness involved. The user uploads the paper and doesn’t have to worry about identifying concepts or types of relationships; they just want “more like this”.
Interestingly, this example, in a presentation ostensibly about knowledge graphs, did not include any knowledge graph, just a list of related articles. Actually, I don’t think there any need for a knowledge graph to be displayed – but that’s for another post.
To summarize, while I completely agree that taxonomies have great limitations, and concepts, working rhizome-fashion, are a great improvement, I don’t think a visual display of knowledge graphs represents the perfect answer to our problems of discoverability. I think the JSTOR Text Analyzer is an interesting experiment in using concept matching, but nowhere near usable as a research tool.
But all credit to Kasenchak and Lehnert for identifying the analogy between postmodernist thought and knowledge graphs. And, just to remind you of the French origin of the analogy, the presentation featured arresting and challenging quotes from the Deleuze and Guattari book, interspersed at various points in the presentation, such as:
- ARBORESCENCE: We should stop believing in trees, roots and radicles. They’ve made us suffer too much.
Now I know it is the trees that were to blame. I’ll never look at a traditional taxonomy (or at a tree) in the same way.