Can an MRS have underspecified predicates?

If I am trying to generate from MRSs, is there a way to say “I want this predicate to be of this type, but I don’t care what the actual word is.”

For example, if I want to be able to generate quantified phrases, but I want a set of different quantifiers to show up in my results (e.g. “a phone,” “the phone”), is there a way to do this if I am working strictly with MRS?

I basically want an EP that’s “more general” than something like the_q so that it would be able to generate multiple results, but this might not exist.

Surely @Dan is the right person to answer about it. Papers about the LOGON translation system can also provide some clues. The @goodmami thesis may have something too. I remember the open-proof project (, it is about generating NL from logical formulas, and it may have in the scripts some ideas about how to start from an abstract MRS and populate that with enough information to be used to NL generation.

Does it make sense?

For some preds, yes, there are little hierarchies! What you want is the SEM-I (semantic interface), which should be erg.smi (I think). It will show you what the possible PREDs are and which ones are underspecifications of which other ones.

1 Like

Yes, @ebender is right that the ERG provides a few small hierarchies of semantic predicates, to let you underspecify some predicates in an MRS for generation. In the 2020 version of the ERG, there is an abstract predicate existential_q' that subsumes _the_q" , _a_q' and udef_q’ so the generator can produce multiple outputs. To try this using the 2020 version (alas not 2018 or earlier), parse “the cat slept” and then in the resulting MRS replace _the_q with existential_q and give this MRS to the generator, to produce both “The cat slept.” and “A cat slept.”. Similarly for “the cats slept”, where the generator will produce both “The cats slept” and “Cats slept.” You’ll find these abstract predicate hierarchies defined not in erg.smi but in the file hierarchy.smi (in the `etc’ subdirectory of the ERG source directory).

1 Like

More information about defining SEM-Is is at SemiRfc · delph-in/docs Wiki · GitHub. There’s also some discussion in my dissertation in Section 2.5.