8.2 Low-Confidence Articles

What about the other direction? Which articles is the model most uncertain about. This is a bit more of a stress test of the model. The high-confidence articles all look pretty much right for the topics they are in. (Not least because I named the topics after the high-confidence articles.) But if the model throws up its hands at articles that are easy to place, that’s relatively bad. So let’s look.

There are more or less sophisticated ways to measure how unsure the model is about an article. I’m going to go with one of the less sophisticated ways, because it is easy to understand and provides clear enough guidance. Implicitly in the previous section, I measured the model’s certainty about an article by the maximal probability it gives to the article being in any one topic. I’ll say it is most uncertain about an article if that maximal probability (for that article) is lowest.

By that measure, here are the ten articles the model is most unsure about.

Table 8.6: Articles the model is most uncertain about.
Alan Donagan, 1970, “The Encyclopedia of Philosophy,” Philosophical Review 79:83–138.
Jay L. Garfield, 1989, “The Myth of Jones and the Mirror of Nature: Reflections on Introspection,” Philosophy and Phenomenological Research 50:1–26.
J. H. Kultgen, 1956, “Operations and Events in Russell’s Empiricism,” Journal of Philosophy 53:157–67.
Harry Prosch, 1972, “Polanyi’s Ethics,” Ethics 82:91–113.
C. J. Ducasse, 1936, “Mr. Collingwood on Philosophical Method,” Journal of Philosophy 33:95–106.
Wilfrid Sellars, 1973, “Actions and Events,” Noûs 7:179–202.
R. Harré, 1973, “Surrogates for Necessity,” Mind 82:358–80.
Donald W. Rogers, 1947, “Philosophic Method,” Philosophical Review 56:656–69.
R. Edgley, 1956, “Critical Notice,” Mind 65:551–7.
Roy Wood Sellars, 1941, “A Correspondence Theory of Truth,” Journal of Philosophy 38:645–54.

Did the model get it right? Should it be uncertain about these articles? Let’s look at some cases, starting with the one it is most uncertain about.

Table 8.7: Alan Donagan, “The Encyclopedia Of Philosophy.”
Subject Probability
Other history 0.0796
Analytic/synthetic 0.0675
Meaning and use 0.0660
Faith and theism 0.0439
Truth 0.0437
Radical translation 0.0383
Physicalism 0.0379
Promises and imperatives 0.0336
Intention 0.0327
Propositions and implications 0.0310
Arguments 0.0309
Universals and particulars 0.0303
Definitions 0.0279
Verification 0.0272
Early modern 0.0236
Ordinary language 0.0234
Denoting 0.0227
Causation 0.0224
Ancient 0.0204

For reasons best known to them, the editors of the Philosophical Review commissioned a critical notice (Donagan 1970) of the eight-volume Encyclopedia of Philosophy. It could be called a book review I guess, but it’s fifty-six pages long, so it feels like it should be in our study.

And I’m fairly happy that this was the article the model had the greatest trouble with. How could it classify a critical notice of an encyclopedia. What topic could it not be in? The answer seems to be the very late topics—there is no wide content or quantum physics there—and the very early topics. I’m actually a little surprised that idealism doesn’t turn up.

So far so good—the model threw up its hands exactly when it should have done so. It would have been wrong to confidently place Donagan’s article. What about the second one?

Table 8.8: Jay L. Garfield, “The Myth Of Jones And The Mirror Of Nature: Reflections On Introspection.”
Subject Probability
Theories and realism 0.0837
Self-consciousness 0.0660
Justification 0.0643
Wide content 0.0626
Radical translation 0.0598
Perception 0.0559
Speech acts 0.0509
Minds and machines 0.0505
Arguments 0.0478
History and culture 0.0445
Physicalism 0.0419
Concepts 0.0366
Frankfurt cases 0.0328
Cognitive science 0.0303
Mechanisms 0.0274
Meaning and use 0.0222
Norms 0.0219
Psychology 0.0211

And this is a bit more depressing. Garfield’s article (1989) is wide-ranging, covering a number of big questions at the heart of philosophy of mind and epistemology. But still, it isn’t that hard to say what it’s about broadly. And to be sure, the model does recognize that this isn’t a philosophy of physics article, or a political philosophy article, or an ancient philosophy article. But still, it should do better than this.

What’s happened, in a picturesque sense, is this: the articles are arranged in a feature space. The feature space has many, many dimensions, and the articles do form clusters within it. What the model does is pick ninety points in that space such that as many articles as possible are reasonably close to one of the points. Now some articles, like the Donagan, are going to be so idiosyncratic that they aren’t near a point. But the Garfield isn’t like that. The model could have decided that Sellarsian theories of mind/epistemology are a focal point. It just didn’t do that. Instead, articles like these ended up falling between many many different points.

I think it isn’t a coincidence that there is another article by Wilfred Sellars on the list. I think it is a coincidence that there is an article by his father though. That one is weird. Here is its table.

Table 8.9: Roy Wood Sellars, “A Correspondence Theory Of Truth.”
Subject Probability
Dewey and pragmatism 0.0972
Propositions and implications 0.0909
Knowledge 0.0841
Theories and realism 0.0784
Truth 0.0777
Perception 0.0739
Psychology 0.0730
Definitions 0.0656
Ordinary language 0.0530
Universals and particulars 0.0514
Idealism 0.0371
Vagueness 0.0354
Analytic/synthetic 0.0303
Self-consciousness 0.0254
Concepts 0.0248
Theory testing 0.0218

In this paper (R. W. Sellars 1941), Sellars is operating at a fairly high degree of abstraction, and considering the ways in which the big philosophical views (idealism, pragmatism, realism, etc) have characteristic theories of truth. The theory of truth he wants to offer is a correspondence theory that isn’t so closely tied to general forms of realism. And we can see why the paper looks, to the model, like it could be about all sorts of things. I’m still a bit surprised that the model didn’t just lump it in with other works on truth. It doesn’t mention the paradoxes, and has no formalism, and maybe that was enough. But it’s surprising.

Anyway, I’m pleased that the article it was most uncertain about was a really impossible-to-place article, disappointed that it couldn’t do a better job with Sellarsian philosophy, and not surprised that it also threw up its hands at various methodology articles. It’s not a perfect model, but it did fairly well.