Skip to content

FeforDroppedArguments

anonymous edited this page Oct 9, 2011 · 8 revisions

Notes from the discussion on dropped arguments: To include or not include pronoun rels? Is it sufficient to constrain the u variable with e.g., png or definiteness information?

Emily's slides (pdf)

The 'eat' example can be misleading, since then we have to worry about whether there's two separate relations. 'earlier' as in "The earlier train" vs. "*The before train" is a clearer case. (Both are relation, take a 'than' complement, but with earlier it's optional.)

"Move!" shouldn't be able to mean "Nobody move!"

Current assumption is that the arity of a given predicate is fixed and known, so leaving off arg roles is not a legitimate means of underspecification.

This leaves:

{{{[ _eat_v_1_rel

  • ARG1 x ARG2 u ]}}}

or

{{{[ _eat_v_2_rel

  • ARG1 x ]}}}

But cf 'earlier' examples where it seems less plausibile to posit two separate relations.

What about dropped handle arguments? ("I persuaded Sandy") If that's just a u, but MRS wellformedness requires a handle to fill that hole, what to do? Possibility of u_ind and u_hand.

If we underspecify the dropped argument completely (leave off its arg_rel), then we potentially lose information coming from the morphology of the verb (e.g., png) or its lexical properties (definite/indefinite null instantiation).

Slippery slope: What about cognate objects? We have "The dog barked an enormous bark" "sleep a long sleep" "die a terrible death". Surely wedon't want to posit ARG positions for the cognate argument that are underspecified everywhere the verb is used without them. Similarly, what about denominal verbs ("to sunscreen someone") --- should the representation have an ARG role for the sunscreen?

Spent some time looking for cases where the verb or other constraints provides information about u that is 'genuinely semantic':

{{{Bought a book, did he? Cut well, don't they? (scissors) Cut well, doesn't it? (scissors)}}}

... relationship to other kinds of anaphora that seem sensitive to grammatical properties of other words for their referents, even if the other words haven't been used in the local discourse context. Also French "le machin/la machine" --- translations of 'thingamajig' with grammatical gender, often used to match the grammatical gender of words that the speaker can't retrieve.

We want to have pron_rel in the cases where there is an overt pronoun because that information is important for coreference resolution. On the other hand, there is already a feature PRON-TYPE which can be used to encode whether or not there was an overt pronoun, and if so, of which type. At the moment, there is a close relationship between pronoun variables and unique EPs. This raises the worry about duplicating the work done by handles with variables, but at least some approaches to information structure require us to talk about relations , not variables. For example, focusing big in big dog is different from focusing dog. (But: they have the same LBL, so how do we get our hands on just one of the relations?)

This raises the question: Does the label of a pronoun_n_rel ever get identified with anything else, or is it always only qeq the RESTR of a pronoun_q_rel? If EPs are what is focused or topicalized, then you can't focus zero pronouns (assuming they don't introduce any EPs). If there is indeed such a distinction between overt and zero pronouns (in some languages at least), do we need the contrast between introducing and not introducing the pronoun_n_rel to capture it? Will PRON-TYPE be sufficient?

What about pronoun resolution in languages which have pronoun-incorporation? (E.g., Bantu but also Romance) Do we require pronoun_n_rel to be introduced by the verbal affixes in order to do pronoun resolution? What about languages like Japanese that have no morphological reflex of pro-drop?

A final worry: How does this affect the MRS to RMRS conversion?

See also FeforPng.

Clone this wiki locally