CauseNet: Towards a causality graph extracted from the web

Random behavior in inputs, or in operations, results in random behavior in the outputs. But there is no statistical expression or characterization that can predict the distribution of one from the other.

You can't say, I want this much distribution in the outputs, so I will add this much distribution to the inputs, weights or other operational details.

Even if you create an exhaustive profile of "temperature" and output distributions across the training set, it will only be true for exactly that training set, on exactly that model, for exactly those random conditions. And will vary significantly and unpredictably across subsets of that data, or new data, and different random numbers injected (even with the same random distribution!).

Statistics are a very specific way to represent a very narrow kind of variation, or for a system to produce variation. But lots of systems with variation, such as complex chaotic systems, or complex nonlinear systems (as in neural models!) can defy robust or meaningful statistical representations or analysis.

(Another way to put this, is you can measure logical properties about any system. Such as if an output is greater than some threshold, or if two outputs are equal. The logical measurements can be useful, but that doesn't mean it is a logical system.

Any system with any kind of variation can have potentially useful statistical type measurements done on it. Any deterministic system can have randomness injected to create randomly varying output. But neither of those situations and measurements makes the system a statistically based system.)

> Ontologies and all that have been tried and have always been found to be too brittle.

I'd invite you to look at ontologies as nothing more than representations of things we know in some text-based format. If you've ever written an if statement, used OOP, trained a decision tree, or sketched an ER diagram, you've also represented known things in a particular text-based format.

We probably can agree that all these things are ubiquitous and provide value. It's just that those representations are not serialized as OWL/RDF, claim less about being accurate models of real-world things, and are often coupled with other things (i.e., functions).

This may seem reductionist in the sense of "we're all made of atoms", but I think it's important to understand why ontologies as a concept stick: they provide atomic components for expressing any knowledge in a dedicated place, and reasoning about it. Maybe the serializations, engines, results or creators suck, or maybe codebase + database is enough for most needs, but it's hard to not see the value of having some deterministic knowledge about a domain.

If you take _ontology_ to mean OWL/RDF, this paper wouldn't qualify, so I'm assuming you took the broader meaning (i.e., _semantic triples_).

> Take the examples from the front page (which I expect to be among the best in their set)

Most scientific work will be in-progress, not WordNet-level (which also needs a lot of funding to get there). You ideally want to show a very simple example, and then provide representative examples that signal the level of quality that other contributors/scientists can expect.

Here, they're explicit about creating triples of whatever causal statements they found on Wikipedia. I wouldn't expect it to be immediately useful to me, unless I dedicate time to prune and iron out things of interest.

> human_activity => climate change. Those are such a broad concepts that it's practically useless.

Disagree. If you had one metric that aggregated different measurements of climate change-inducing human activity, and one metric that did the same for climate change, you could create some predictions about N-order effects from climate change. Statistical analysis anyway requires you to make an assumption about the causal relationship behind what you're investigating.

So, if this the level of detail you need, this helps you potentially find new hypotheses just based on Nth order causal relations in Wikipedia text. It's also valuable to show where there is not enough detail.

> Or disease => death. There's no nuance at all.

Aside from my point above - haven't looked at the source data, but I doubt it stops at that level. But even if it does, it's 11 million things with provenance you can play with or add detail to.

Or you can also show that your method or choice of source data gets more conceptual/causal detail out of Wikipedia, or that their approach isn't replicable, or that they did a bad job, etc. These are all very useful contributions.

Essentially one totalizing meta analysis.

E.g. If it reads an article about the relationship between height and various life outcomes in Indonesian men, then first, it would store the average height of Indonesian men, the relationship between the average height of Indonesian men and each life outcome in Indonesian men, the type of relationship (e.g. Pearson's correlation), the relationship values (r value), etc. It would store the entity, the relationship, the relationship values, and the doi source.

Something like a quantitative Wikipedia.

On the other hand, what we seem to have with LLM models, and the transformer approach in particular, is a sort of probable statistical correlation, calculated by brute-forcing and approximation (the gradient descent). So this is not true causation also, it becomes one only after a human observes it and agrees it follows certain causality.

/Not sure whether I can state that it is also material but in another non-logical sense, perhaps would sound nonsensical, but the apparent logical structure in the LLM production rather emerges from training patterns, not from explicit logical operations./

There's nothing wrong having a graphical structure which models causality, and of course - this needs to be discovered first. But then we have LZW/Sequitur using very brute-force way in order to find the minimal grammar for compressing certain data lossless-ly, thus discovering some logical structure (and correlation), but this is not yet causation. Indeed finding patterns != finding causal relationships.

My gut feeling is we want something that would result in a correct PROLOG-like set of inference rules, but based on actual causality, not conflating correlation. And then this - for a larger corpus - world's knowledge, but we don't have the means (yet) to figure out the correlation, even though approaches exist for smaller corpus.

It is perhaps the gradient descent and the fact that this composition of tensor algebra is differentiable that is the ingenious thing about the ML we deal with now, but everyone is dreaming of some magic algo which would allow finding the causation so that it results in non-probabilistic graphical model, or at least a model that we can follow the stochastic branching on in a observable manner.

It is indeed ingenious to fold multi-dimensional spaces, multiple times, in order to disambiguate the curvature of bunny's ear from the one of bear's ear. But it just does not feel right to do logic and causation by means of differential calculus and stochastic structures.