Looking at their Neo4j graph, they also retain the provenance of the causal relation in "claimedIn" relations between the reified triple of each cause-effect pair. So, that's at least marginally useful for fact-checking or quality evals.
> they just take noun phrases and slap an underscore between the words and call it a concept.
Not to defend lazy approaches, but you could make this point about tokens also ("take any bunch of characters that happens often enough, and call it a token").
> It'll only work for highly specialized, unambiguous concepts
Fair point. Practically, there's not much use in this unless you really dedicate time to figure out what's meant by each concept, and prune junk. And by that time, you may as well make your own pairs.
But from a research POV, hardly anyone will go through such effort, so I still find it quite useful. Some potential questions that I could derive from this (aside from ~110 works citing it since 2020, which is not bad for KRR work):
- What is the quality of causal relations (e.g., diagnostic decision trees) in medical articles?
- How can the original scientific provenance of cause-effect pairs best be represented?
- Which extracted causes are true variables in some effect, and to what extent/direction? (i.e., an ablation study at scale, provided you can find appropriate data)
I think we look at it from different sides. Mine is "how good will this be in independently running code" (where 80% correctness is minimally needed), yours seems to be more "how well does this represent our knowledge" (from different angles).
> you could make this point about tokens also ("take any bunch of characters that happens often enough, and call it a token").
My reason was more that such an approach doesn't work well with (unrestricted natural language) text. E.g. side_effects => physical_damage: What side effects? (Why plural?) Not all cause physical damage. And not all side effects that cause physical damage cause the same damage. The differences are described elsewhere in the text, but not consistent enough to extend the token with that information, so just associating literal excerpts from a text will practically guarantee underspecification (except for practically unambiguous term). The effectiveness will be language dependent, of course.
But from a research POV, hardly anyone will go through such effort, so I still find it quite useful. Some potential questions that I could derive from this (aside from ~110 works citing it since 2020, which is not bad for KRR work):
- What is the quality of causal relations (e.g., diagnostic decision trees) in medical articles?
- How can the original scientific provenance of cause-effect pairs best be represented?
- Which extracted causes are true variables in some effect, and to what extent/direction? (i.e., an ablation study at scale, provided you can find appropriate data)