RmrsDesign

TableOfContents

Sets of interests or constituencies that have a stake in the design of (R)MRS representations:

• parsing/underspecification
• generation
• integration with shallow systems
• shallow inference/IE
• inference
• syntax, composition
• multilingual harmonization
• transfer
• anaphora

Agenda items

• cross-linguistic similarity
• paraphrase
• robustness, underspecifiability
  • predicate names
• stability, documentation (core phenomena)
• abstract predicates
• decomposition
• near-equivalences
• clean IE
• sense mapping
• `real' documentation
• naming conventions for unknown word predicates
  • proper names

Demise of pronoun_q_rel:

When we remove pronoun_q_rel, we run into the problem that the tree gets smaller, and the labels of the pronoun relation have to get identified with something. This was troubling for quite a while until we convinced ourselves that modified pronouns are always pseudopartitives.

Pronouns don't ever get quantifiers, and share their labels with the verbal projection they combine with ("You who brought laptops probably have the right idea." "You with laptops must leave now."). Modified pronouns ("You who brought laptops") involve a pseudopartitive construction which introduces both a new variable and a quantifier for that variable.

Proper names can go without quantifiers, but might get them when they are overt ("Every Kim I've met"). "John in Paris" doesn't need a quantifier.

Optionality of pronoun_n_rel

In languages with pro-drop (of various kinds), dropped arguments often fill the role of unstressed pronouns in a language like English (used for reference to discourse entities currently in focus). But dropped arguments can also have other uses (e.g., like indefinite null instantiation in English). The question is what would break if we didn't put in any ep for these dropped arguments.

Raises questions about the notion of characteristic variables/eps: Every ep has an ARG0 and every non-u index is the ARG0 of exactly one ep (with special exceptions). The rmrs matching code uses this notion, and drops those variables that can't be grounded in an ep for which they are the characteristic variable. The notion of characteristic variable is also relevant to dependency extraction code (used e.g., in producing n-gram models against which to rank transfer outputs) and potentially to filtering lexical edges based on information from other lexical edges in the gender.

In terms of the model theory, both sleep(e,u) and sleep(e,x) are fine, but the lack of any characteristic ep affects the complexity of tasks such as deciding the mutual satisfiability of two rmrs.

PRED with internal structure

It may be useful to make explicit the internal structure of (open-class) PRED values, which by common consent are currently strings which have the following structure: "_LEMMA_POS_SENSE"

We explored a proposal to change PRED so its value is no longer a string for open-class lexical entries, but a type with the following structure:

• [ LEMMA string, POS type, SENSE type ]

There are several possible benefits of this increased transparency:

• More direct comparison of MRS and RMRS structures

• Enabling of lexical redundancy rules like causative-inchoative in English, where each of the entries has a single EP, and where the PRED value in one entry is systematically related to that of the other entry. For example, with "open", the ERG has two entries, the intransitive verb with "_open_v_1", and the transitive verb wtih "_open_v_caus". On this proposal, the redundancy rule would replace just the SENSE value '1' with 'caus'. (Note that the LKB does not yet have TDL notation defined to support a definition of a lexentry1 as lexentry2+lexrule3, but this is not expected to be hard. Note also that such a redundancy rule is a descriptive device, not a unary rule to be applied in parsing or generation, since it does not preserve monotonicity of the MRS. (But it was suggested that this monotonicity could be preserved if the entry for 'open' were underspecified for SENSE, with two rules, one for inchoative and one for causative.))

• Improving treatment of subregularities for verb-particles, such as with semi-productive "up" as in "wake up". Here it is good to have SENSE as a type, since we could arrange its value here to be something like the following, in order to capture the causative-inchoative alternation for "wake up". up caus

  • \ /

  up-caus

Non-restrictive relative clauses

The following MRS is proposed for non-restrictive relatives, making them analogous to appositions. For this example:

Dogs, which bark, sleep.

the target MRS will be roughly as follows, introducing a distinguished quantifier (ignoring HCONS here):

l1:dog_n(x1)

l2:udef_q(x1,h1,h2)

l1:appos(x1,x2)

l3:nonrestr_q(x2,h3,h4)

l4:generic_nom(x2)

l4:bark_v(e1,x2)

l5:sleep_v(e2,x1)

Predicate Hierarchies

It would be interesting to add external hierarchies (for descriptions or denotations) to the current predicates. However, this is a non-trivial extension to the current software.

These hierarchies would have to be made available to MRS manipulation routines such as MRS comparison and generation.

Currently at last two kinds of groupings have been proposed:

(a) subsumption relations

• (a-1) subsumption relations with fixed arity (can be used to underspecify generation)

(b) lemma-pos-sense underspecification

Such hierarchies could be description level (bottoming out in MRSs) or denotation level (bearing a relationship to ontological truth within some domain). They would also be application/domain/MT-pair dependent, which particular applications calling the appropriate hierarchies at appropriate points.

FCB volunteers to cheerfully test this in MT.

EPs v. properties

How do we decide for any piece of information we would like to encode in the MRS whether it should be a variable property or an EP? Some guidelines:

1. If it is constrained by multiple things in the syntax, it is more convenient

for it to be a feature. It may be possible to specialize PRED values from multiple sources, but that might also violate compositionality.

1. If it takes more than one argument, it should be an EP.
2. Quantifiers have to be EPs, even though compositionally they only take one

argument (never directly associating their BODY element with anything), because it's not enough to know which variable they are associated with: In "All brown dogs bark", there are 'x's appearing in brown, dog, and bark, and we need to know which belong in the definition of which of the sets that the quantifier is comparing (RSTR/BODY).