I recently had an email conversation on the topic of making AMR more formal in a set-theoretic sense. Expanding on it here as it seems relevant to our future-of-AMR conversations:
While I'm not really an expert on set-theoretic semantics, I think we would have difficulty coming up with guidelines for such semantics in a sentence like:
"Not all of the passengers were warned of the railway closures due to tunnel construction."
Unless it is explained in the context, I don't actually know if it is one or more tunnels being constructed; whether it should be considered one overall construction event, or several; whether the whole railway is closed, or just part of it, or different parts at different times; what is meant by "the passengers" (all people who might be affected? a specific group of people mentioned previously in the discourse?); if it is multiple closure events, whether different subsets of passengers were warned about different closures; and how the negation should be represented ('it is not the case that all passengers were warned...' or 'some but not all passengers were warned'?).
At a deeper level, this sentence seems to suggest that warning all of the passengers would have been appropriate/expected—so a pragmatically-sensitive analysis might be something like NOT(w) INSTEAD_OF_EXPECTED w, where w stand for the proposition "all passengers were warned...".
I think it is probably unrealistic to expect annotators to consider such issues systematically when creating the original AMR. But a couple of alternatives might be viable:
We prescribe special treatment only for a clearly circumscribed set of phenomena. E.g., "generic constructions" (declarative statements about nonspecific plural NPs, as in "Cats are cute") would have a special marking in the AMR, while most NPs would not.
We make this a second-pass discourse-level annotation task, as suggested in #170. It would likely be a more difficult task than creating the original AMRs, as annotators would have to understand the AMRs as well as the discourse annotation scheme.
I recently had an email conversation on the topic of making AMR more formal in a set-theoretic sense. Expanding on it here as it seems relevant to our future-of-AMR conversations:
While I'm not really an expert on set-theoretic semantics, I think we would have difficulty coming up with guidelines for such semantics in a sentence like:
"Not all of the passengers were warned of the railway closures due to tunnel construction."
Unless it is explained in the context, I don't actually know if it is one or more tunnels being constructed; whether it should be considered one overall construction event, or several; whether the whole railway is closed, or just part of it, or different parts at different times; what is meant by "the passengers" (all people who might be affected? a specific group of people mentioned previously in the discourse?); if it is multiple closure events, whether different subsets of passengers were warned about different closures; and how the negation should be represented ('it is not the case that all passengers were warned...' or 'some but not all passengers were warned'?).
At a deeper level, this sentence seems to suggest that warning all of the passengers would have been appropriate/expected—so a pragmatically-sensitive analysis might be something like NOT(w) INSTEAD_OF_EXPECTED w, where w stand for the proposition "all passengers were warned...".
I think it is probably unrealistic to expect annotators to consider such issues systematically when creating the original AMR. But a couple of alternatives might be viable: