EdsConversion

Background

Basic Procedure

Disambiguation Heuristics

Up until the 1214 release of the ERG at least, there are some predications that encode aspects of information structure rather than core predicate–argument relations. These include the ‘discourse’ relations introduced by the grammatical constructions of passivization and topicalization, as well as two-place relations that effectively express an identity relation between two (distinct) instance variables. The latter class includes the appos(ition) relation and the id(entity) relation (used in tag questions and some coordinate structures). These relations are generally dis-preferred (and are likely to be recast in terms of ‘indvidual constraints’ in forthcoming versions of the MRS framework), and there is a grammar-specific ‘black list’ of these predicate names (called *eds-non-representatives* in the Lisp implementation).

The most common cause of one-to-many correspondences between a variable and a set of predications are labels shared with (non-scopal) modifiers, e.g. in a structure like she arrived very quickly. Here, the degree specifier is a non-scopla modifier on the adverb, which in turn is a non-scopal modifier on the arriving event; thus, all three share one label, and arrive is the ARG1 of quickly, which is the ARG1 of very. To pick out arrive in this scenario, we dis-prefer candidates that take any of the other candidates as their argument. This is a sound topological heuristic, essentially operationalizing a notion of semantic heads in groups of (logically) conjoined predications.

Far less frequent than the above are cases of two or more predications sharing their label but lacking argument relations among them. In the 1214 release of the ERG, the ‘existential be’ constructions can give rise to such configuration, e.g. in there were cats in the garden (mrs/991). Here, the preposition shares its label with the _be_v_there, but its external argument (ARG1) is the cats instance variable. Similar configurations can arise with ‘it clefts’, e.g. it was Browne whose manager interviewed Abrams (csli/977). Here, the proposition embedded by the _be_v_itcleft relation shares its label with the two-place poss(essor) relation holding between Browne and his manager. To disambiguate cases like these, there is a dis-preference for relations whose intrinsic variable is ‘untensed’ (where a grammar-specific parameter—*eds-untensed* in the Lisp implementation—provides an appropriate test, as a pair of a variable property and ‘untensed’ value).

Predicate Modification

The EdsTop page discusses the somewhat common instances of ‘fragmented’ EDS graphs owing to the incomplete ERG analysis of degree specification on quantifiers, as for example nearly every. In no small part to address these imperfections, conversion to EDS provides an optional, semi-declarative mechanism akin to semantic predicate modification, i.e. mapping the actual ERG analysis of the above to something that one might interpret as nearly(every).

Predicate modification is triggered by certain classes of predicates who ‘lack’ an argument, i.e. conceptually the triggers are pairs of a predicate pattern and an argument label, for example something like 〈/_x_deg$|^_quite_x$ /, ARG1〉 for the ERG (in the Lisp implementation the trigger patterns are configured as *eds-predicate-modifiers*, while in mid-2016 the corresponding argument labels are sadly hard-wired as ARG1). When there is an MRS predication whose predicate matches one of the triggers, the conversion procedure (a) checks that the corresponding argument (e.g. ARG1) is not present or unbound (of type u), and that (b) there is at least one other predication sharing the same label. Where this is the case, the argument will be specified to become that label (in a sense creating what would look like a self-referential structure if intepreted as an MRS), such that the EDS node corresponding to the degree specifier (e.g. nearly) will end up with an argument edge pointing to the node of the quantifier (all, say).