If you build anything that classifies persuasive language, you will at some point sit down to enumerate the techniques. This seems like a bounded task. Aristotle started it, the medieval logicians extended it, and there are published lists.
It is not a bounded task, and the reasons why turn out to be interesting for anyone doing classification work in domains that predate computing.
The categories overlap, and they overlap essentially
Take three well-known devices: the false dilemma, the loaded question, and the motte-and-bailey.
A false dilemma presents two options as exhaustive. A loaded question embeds a premise the answer must accept. A motte-and-bailey advances a strong claim and retreats to a weak one under challenge.
Now consider: "Do you support the reform, or do you want the system to collapse?"
That is a false dilemma. It is also a loaded question — answering either way concedes that collapse is the alternative. And if the speaker retreats under pressure to "I only meant we need some reform", it is also a motte-and-bailey.
One sentence, three categories, and no principled way to pick one. This is not an edge case; it is typical. Rhetorical devices were catalogued by different people for different purposes over two millennia, and nobody was maintaining a schema.
Multi-label does not fix it
The obvious engineering response is multi-label classification. Let the sentence be all three.
This helps with the immediate problem and creates a worse one downstream: the labels are no longer independent, and anything you compute over them inherits the correlation structure. Count occurrences and you double-count. Aggregate into a summary and you weight the overlapping regions more heavily for no principled reason.
You also lose the ability to say anything meaningful about frequency. "The false dilemma appears in 15% of articles" means something quite different if half those instances are also counted as three other things.
Prototype categories, not definitions
The deeper issue is that rhetorical categories are not defined by necessary and sufficient conditions. They are prototypes — a central clear case, and a fuzzy boundary of things that resemble it.
The central case of an ad hominem is obvious. The boundary is not: pointing out that a source has a financial interest in their claim is either a relevant observation about reliability or an attack on the person, depending on whether the claim is being evaluated on evidence or on testimony. The same sentence can be both, depending on the surrounding argument.
Classification systems built on necessary-and-sufficient definitions handle prototype categories badly. They produce confident answers on the clear cases, which is where you did not need them, and arbitrary answers on the boundary, which is where you did.
What I would tell someone starting this
Three things I wish I had internalised earlier.
The catalogue is not the model. Published lists of fallacies are pedagogical artefacts, organised for teaching, not for computation. Taking one and treating it as a schema imports two thousand years of unmanaged category drift.
Frequency claims need a stated counting rule. "How often does technique X occur" is unanswerable without deciding what to do about overlaps, and the decision changes the answer materially. Any published frequency should come with the rule attached.
Boundary cases are the product. The clear cases are handled by a reader's ordinary competence. The value of any tool here is in the middle, which is exactly where the taxonomy is least reliable. Designing as if the clear cases were the target produces a system that is confident where it is useless.
The consolation
Working through this, one thing became easier rather than harder.
Most of the techniques, whatever their traditional names, do one of a very small number of things: they shift what is being evaluated, they set a frame before evaluation begins, or they exploit the asymmetry between the cost of making a claim and the cost of checking it.
The named devices are combinations and special cases of those moves. Which is encouraging for a reader — you do not need forty reflexes, you need three questions — and it suggests that the useful computational unit might not be the traditional category at all.
I am not confident about the three. It feels suspiciously tidy, and tidy taxonomies of messy phenomena are usually compressing something that will matter later.
We publish individual analyses that show where the boundaries fall in practice at faktoskop.pl — including the cases where the classification is genuinely arguable, which are the informative ones.
Question for the community
For anyone who has built classification over a domain with a long pre-computational history — legal categories, medical taxonomies, linguistic frameworks: how do you handle inherited category systems that were never designed to be disjoint?
My instinct is that you have to build your own schema and accept losing comparability with the literature. But that loses a lot, and I would like to hear from people who found a better trade.
Top comments (0)