Differences between MRS and SEMENT

For my project I’ve been working on implementing the algebra from the MRS algebra paper. As I’m writing about it, I’m trying to understand the differences between the definition of a SEMENT as discussed in that paper vs. the formal definition of an MRS from the 2005 paper.

From what’s stated in the papers, an MRS is a 3-tuple <GT, R, C>:

  • GT – global top
  • R – bag of EPs
  • C – handle constraints

Whereas a SEMENT is a 5-tuple <HK, H, R, EQ, C> (I made up these letters):

  • HK – hook, has an index and a top handle, so similar to GT but with an index
  • H – set of holes
  • R – bag of EPs
  • EQ – list of eqs
  • C – handle constraints

So SEMENT has two additional elements, and the hook has both an index and a top handle. So what I’m trying to understand is the reason for each of these differences.

Here are my guesses:

  1. As far as the MRS not having an index I don’t really have any guesses, except that @ebender and I discussed it and she thinks it has to do with the MRS structures as described in the paper being “complete” so there isn’t a need for an index anymore, but during composition I would think an index is necessary, and also the MRSs that come out of the ERG still have an index so I don’t know what to make of that
  2. A SEMENT needs a list of EQs to keep track of which variables are equated. An MRS doesn’t need EQs because composition is done through unification, so if some grammar rule says that the index of a constituent and the index of its argument are equated, the unification process will resolve this so you don’t need to keep track of it in a list. (this is how @ebender described it to me but I might be messing up the details)
  3. A SEMENT needs a list of holes so that it’s clear what arguments need to be plugged. As far as an MRS (during composition) my guess is that the lexical entries do the work of specifying how syntactic arguments map to semantic roles. e.g. transitive verbs will state that (in active usage) the subject gets mapped to ARG1 and the object to ARG2. So maybe this is why keeping track of holes isn’t necessary? But if you’re doing it purely algebraically without lexical entries you have to keep track of what holes are available and whether they’ve been filled?

Any help would be appreciated! Thanks!

  1. That’s correct. The INDEX is irrelevant for the semantics of a complete MRS viewed as an underspecified logical structure. It was, however, convenient for some users to retain the INDEX in the complete MRS as a way of identifying some subparts of the structure, so that’s why it’s output for the complete MRS from the ERG and other grammars.
  2. Yes, essentially. One could, of course, describe a complete MRS in terms of distinct identifiers with equalities between them, but this would be cumbersome. When we’re talking about the SEMENTs, we wanted to describe the process of composition, which is essentailly all about creation of these equalities, so that’s why they are explicit.
  3. Yes, essentially. The algebra abstracts away from the individual grammars and indeed the implementation via a typed feature structure formalism. This is made more explicit in https://aclanthology.org/W07-1210.pdf
1 Like

Adding to @AnnC’s comments:

  1. One reason that INDEX on a full MRS might still be useful is that In running text, distinguishing one utterance from another is not always so sharp, particularly with non-standard punctuation: “We arrived yesterday - it was great to be back home.” could be treated as having just one MRS with some kind of implicit conjunction (using the INDEX values supplied by the MRSs for each of the two clauses), or as two separate MRSs as we would for “We arrived yesterday. It was great to be back home.” It is also convenient for humans interpreting an MRS to have the INDEX identifying the scopally outermost EP apart from quantifiers; starting from the global top is more cumbersome since the human has to step through the relevant qeq(s) to find that EP.

  2. Referring to a list of holes is useful in making explicit which semantic composition operations a particular lexical entry (such as a control verb like “try”) uses to combine with its arguments, but these holes are also necessary in describing the semantic effects of grammar rules, some of which introduce additional semantics beyond what their daughter(s) will supply. For example, the unary grammar rule or rules enabling extraction of a modifier (used in e.g. “Where did he say we should [wait __]?”) must attend to whether the modifier phrase that eventually fills that syntactically introduced gap is a scopal one or nonscopal, with markedly different semantic effects.

1 Like

Ah, that second comment should have been labeled as “3.” since it was intended as a response to your third question about holes, not the second one about equalities.

1 Like

I am a little lost about your comment on the holes. For one thing I can’t off the top of my head picture what would be going on with the modifier extraction example. But also, I thought that the MRS structures in the ERG did not keep a list of holes the way that the SEMENT structures in the algebra paper do. But it looks like you’re saying you do (or at least it’s useful?) to keep track of holes.