Copestake on Minimal Recursion Semantics

Ann Copestake, Dan Flickinger, Ivan Sag and Carl Pollard. Minimal Recursion Semantics: An introduction. Journal of Research on Language and Computation, 3(2--3), pages 281-332, 2005.

I have read earlier drafts about MRS, and I find this published paper a lot clearer. Earlier drafts had confusing features called Handels and Lizsts. The new presentation seems better motivated. The first pass presentation is more like predicate logic with generalized quantifiers, but using elementary predications (EPs) as atoms, usually corresponding of a single lexeme. Then this is translated to Typed Feature Structures, before incorporating it into HPSG. The resulting structures can be quite elaborate even for a short sentence like "Every dog probably sleeps," but that is typical of HPSG. If this is psychologically realistic, and realizable in a biological neural network as well as a symbolic program, I think the brain has enough neural pathways to handle the typed structures. This approach steps back from using very specific relations, like CORPSE for a participant in DIE, using "semantically-bleached" ARG1 instead. They do mention the possibility of using more specific semantic (thematic) roles, like the ACTOR and UNDERGOER of Anthony Davis. I may want to come back to this, since Davis develops his theories to cover Austronesian languages I am interested in. As an aside, I wonder what work on thematic roles has been done for Chinese?

The motivation of this approach is to produce "flatter" representations, in the spirit of "shake-and-bake" translation approaches but more precise. An explicit goal is to support semantic transfer approaches to machine translation. MRS structures were originally developed during the VerbMobil project, and may be one of the enduring legacies of that work.

Many of the examples are related to quantifier scoping, the old donkey anaphora problem. I never felt that problem was very exciting, but a lot of analysis has been poured into it. There is some mention of Discourse Representation Theory, but I still don't see how it fits in with MRS. And now there is dynamic semantics and Segmented DRT as well. So much to do, so little time. I have a feeling that the MRS work is more general than just scope resolution, but the details are not yet clear to me.

The paper gives an example of how the MRS approach can be applied more generally, to phrases like "allegedly difficult problem." This section talks about intersective and scopal modifiers. They also claim that MRS solves problems with Cooper storage, which is part of HPSG that I never found convincing.

MRS has been implemented in the English Resource Grammar of the LinGO (Linguistic Grammars Online) project. Copestake has been working on the DELPHIN project, which talks about deep and shallow processing. I would be interested in seeing if MRS could be applied to a Chinese corpus (shallow) and treebank (deep).

I found the paper interesting, and worth rereading, perhaps after reading Copestake's draft on Robust MRS, which is designed for shallow processing. At that time, I may make comments on the references, some of which I would like to follow up on.