Notes on representations of PNG in (R)MRS

Considering transfer, monolingual, and transfer-level underspecification

This topic arose because of an issue at the intersection of transfer (harmonizing (R)MRSs across languages) and within-language efficiency. In many languages, the agreement facts motivate underspecification (e.g., the type "non third-singular" for English subject-verb agreement), but the exact shape of that underspecification varies from language to language. If two languages don't have the same set of underspecified types in their hierarchies, transfer becomes painful. On the other hand, if two languages do have the same potential for underspecification, it would be nice to preserve that rather than redundantly spell out all of the possibilities in the target language.

The proposal we arrived at was to come up with a grammar-external representation for transfer (not necessarily really a type hierarchy) that includes nodes for all of the possible underspecifications (a power set) of person and number combinations. Individual grammars will provide a mapping from their own types into this external representation. The generator should (? already does?) allow for specialization of feature values in order to generate from an underspecified input.

The external resource would only include person and number, because these draw on a relatively small vocabularly in the world's languages (3, maybe 4 persons, up to about 5 number distinctions) and because this information should be preserved in transfer.

This contrasts with gender which generally shouldn't be preserved in transfer. Certainly not for NPs (i.e., grammatical gender), and in the case of pronouns, the right thing is probably to do reference resolution first and then choose the properties of the target language pronoun on the basis of the properties of the referent/antecedent.

However, the solution proposed here should allow gender to be included in the types used for underspecification within languages (e.g., German sie which is 3sg feminine or 3pl) --- the mapping out could 'lose' the gender information. Doing this would require including gender in the MRS, which is perhaps controversial.

The alternative proposal for gender is to include it not as a feature but by subtyping pronoun relations (such that gender in effect only appears in the MRS in pronouns). Thus English would have _he_n_rel, _she_n_rel, _it_n_rel, _they_n_rel (briefly worry about redundant specification of number contrast in feature and type). This information is of value for reference resolution, which operates on (R)MRSs. Using subtyping instead of a feature is meant to suggest that this information is a constraint on the resolution of the pronoun: not a property of the referent but a property of how it's being referred to.

It's clear that gender must be represented as a feature somewhere because many languages show agreement in gender (or more generally, noun class). On the other hand, this could be done with a syntactic AGR feature. Representing (semantic) gender via rel subtyping does put some constraints on analyses of pronoun incorporation in certain cases. Consider the French example:

{{{Je l'ai cuite I 3sg-have cook-perf-fem 'I cooked it (fem sg)'}}}

In this case (assuming we're introducing pronoun relations for pronominal affixes like so-called clitics in French) the pronoun relation will be introduced by the prefix on the finite verb. However, before a vowel, the contrast between le and la neutralizes. The past participle shows feminine agreement. This can be handled by introducing a local ambiguity at l'ai, where there are in fact two constituents, one introducing _le_n_rel and one introducing _la_n_rel. They are also distinguished in their value for the GENDER feature, and cuite constrains the GENDER value of its argument to be feminine. This seems somewhat redundant. (See also FeforDroppedArguments.)

FeforPng (last edited 2011-10-08 21:12:08 by localhost)

(The DELPH-IN infrastructure is hosted at the University of Oslo)