Discussion: MRS Design
Moderator: Laurie Poulson; Scribe: Antske Fokkens
What is the purpose of T&A in the MRS?
- Define the temporal interpretation of the sentence.
- Provide input to an external processor. -- Which processors?
- Provide only that information needed to constrain ungrammaticality.
What should be in the MRS marking T&A?
- Category labels, e.g., past, perfective, accomplishment
- Semantic features, e.g., E, R, S; completive, bounded; durative, telic
- A combination of the two.
Should T&A features be semantic or syntactic or both?
- Does it depend on the language? on the construction?
- What are the issues?
- What determines whether a verb gets it's own predicate? What are the tests?
- Scope, e.g., of negation?
- Additional semantic content, e.g., wanted to eat, finished eating
- Is some interpretation of multiple events required?
* MRS design, focus on tense/aspect
Laurie: we started from a point to construct grammars that are linguistically grounded. Lead to the first purpose question. For instance, the issue of using a predicate or a feature, how to decide?
Berthold: discussion topic coming after: internal or external MRS? In other words: shall we drop efficiency issues in favor of perfect representation?
Stephan: Answer to what are internal and external MRSs: 2 layers of representing semantics: feature structure and the real MRS, which used to be not a one to one corresponding, but now leading to process of extracting the actual MRS. There are grammar internal distinctions, not relevant for the outside, but for several reasons part of the semantics.
Dan: example: it is useful to know whether something is coordinated or not, this is not present in syntax in theory. This is registered in the semantics (grammar internal in distinction), but is not part of 'to show analysis'.
Berthold: interested in thinking about that, because of difference between Hausa and German: aspect oriented versus tense oriented. When doing mapping underspecifications need to be correct.
Laurie: it is hard to find an abstract representation that any language would use: harmonizing will fail, specially with many languages at play. In the end, one will need to use transfer rules.
Dan: Serial-verbs: we should ask Lars and Dorothee.
Berthold: multiple event approach: events are sometimes on items which are not always events (e.g. adjectives)
Laurie: how to do things like serial constructions in MRS: you can not change values in MRS, you can only specify them. There may be cases you may want an event with an associated event, to prevent underspecifying and having many defaults.
B: technically, for generation equivalence: it is easier to have optionals as variants than as predicates. Information structure features, it would be easier to underspecify using features than with relations.
L: phenomena of interest: Russian secondary imperfective: prefective verbs can take an imperfective marker (repeated perfective event), seems like an additional predicate to me.
D: pseudo-criteria: the underspecification criterion is not simple. Now it is easy to get more specific values when generating, but it does not add new things, internal technical motivations are used for making that choice.
B: there are cases where one would like to treat differently marked items systematically as being identical.
D: at the cost of some additional ambiguity, we can retain monotonic semantic composition. it is possible to have both an additional predicate and a no additional predicate interpretation, coming from different sources, which in certain cases will be disambiguated.
L: the concern is more about the form of the MRS (talk about later) Strong belief each language comes in its own way.
L: is there a notion of what we are modeling exactly? How to use features to consistently reflect the same thing (even within a grammar), how to deal with sometimes marking morphology and sometimes marking temporal interpretation
B: We shouldn't beat ourselves up to matching morphology and temporal marking: some tense markings just happen to be a combination of certain forms (this is morphology)
L: the question is what are people expecting from TAM-markings, what information is useful?
D: it is all useful, maybe vital. But inference should be done elsewhere, not in the MRS as unification. I'd be happier with a separate component: take what our grammar gives us, and do these inferences from there. I want to preserve everything that was there on the surface, that is potentially relevant to the interpretation. This is my view today. ... There might be reasonable reasons to add inference, but we don't know that yet.