For the first ten or so years of EDS support, there used to be only one implementation of the conversion procedure, maintained by StephanOepen as part of the Common Lisp MRS library (which is typically loaded as part of the LKB, but is at times also compiled stand-alone, for example to link with the PET parser). Starting around 2014, RebeccaDridan has developed a C++ implementation of basic MRS manipulation, including conversion to EDS; and in 2016 MichaelGoodman is adding EDS support to the pyDelphin libraries. Although the basic conversion procedure (see below) is straightforward, there are some ‘disambiguation heuristics’ involved that need to be synchronized across EDS implementations. The purpose of this page is to spell out in semi-formal terms all the relevant parameters and heuristics that enable a deterministic result when converting a (possibly mildly ill-formed) MRS to an EDS graph, no matter which implementation of the conversion is used.

Basic Procedure

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).

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); for example in Browne, the manager, arrived.

  h:_arrive_v_1[ARG1 x0]
  named[ARG0 x0]
  _manager_n_of[ARG0 x1]
  h:appos[ARG1 x0, ARG2 x1]

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-scopal 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.

  h:_very_x_deg[ARG1 e0]
  h:_quick_a_1[ARG0 e0, ARG1 e1]
  h:_arrive_v_1[ARG0 e1]

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 cat instance variable.

  _cat_n_1[ARG0 x]
  h:_be_v_there[ARG1 x]
  h:_in_p[ARG1 x]

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—called *eds-untensed* in the Lisp implementation—provides an appropriate test, as a pair of a variable property and ‘untensed’ value).

Finally (for the time being), another systematic ambiguous class was identified when comparing EDS conversion across different code bases, viz. non-scopal modification of scopal predications, for example in Browne merely doesn't work. (csli/795).

  h0:neg[ARG1 h1]
  h0:_mere_a_1[ARG1 e]
  h2:_work_v_1[ARG0 e]
  { h1 =q h2 }

For parallelism with, say, Browne doesn't work (and Browne merely works), it is desirable to select neg in the above as the representative node. To accomplish this, an extension (or maybe generalization) of the initial ‘modifier’ heuristic (see above) is applied, where for each of the candidates two sets of predications are determined, viz. (a) the transitive set of scopal arguments and (b) the immediate set of non-scopal arguments (in determining both sets, the standard assumptions for conversion to EDS are applied, viz. selection of one distinguished variable per predication and interpretation of =q handle constraints as identity). For the example above, set (a) contains the _work_v_1 predication for neg and is empty for _mere_a_1; conversely, set (b) is empty for neg and contains the _work_v_1 predication for _mere_a_1. Given these argument relations, candidate representatives are dis-preferred for which there is a non-empty intersection of their set (b) with the set (a) for any of the other candidates.

Open Questions

Systematic comparison across three independent implementations of the conversion procedure to EDS has uncovered a number of interesting ‘corner cases’. As can at times be the case, one might ask whether the underlying ERG analyses are ‘well-formed’ (e.g. in the sense of meeting the algebraic constraints assumed for MRS composition) or whether they are linguistically fully adequate (in the sense of capturing strong generalizations). However, the EDS philosophy tends to aim for robustness to a broad range of MRSs, hence disambiguation heuristics that lead to a deterministic result should be available for any such cases, either way.

One complication is related to non-scopal modification of noun phrase coordination Presumably for algebra-related reasons, the ERG does not have access to the label of the group-forming conjunction relation (which is embedded below a covert quantifier), and hence the modifier ends up sharing its label with whatever predicate takes the group instance variable as its argument. For example, for (a simplification of csli/577) A programmer and an engineer from Berlin were hired.:

  h0:_and_c[ARG0 x]
  h1:_from_p[ARG1 x]
  h1:_hire_v_1[ARG2 x]

Obviously, the hire node should become the top of the corresponding EDS graph, but it is not immediately obvious how to derive this result in terms of observable MRS properties. The Common Lisp EDS conversion has long used to resort to comparing ‘incoming’ and (transitive) ‘outgoing’ link counts to break ties between multiple candidate representative nodes, like in the above. However, while higher degrees of ‘connectivity’ with the graph at large may seem an intuitive notion, links counts may of course still leave remaining ambiguity, and they may also give unwanted results, for example if a non-scopal modifier in a construction like the above had a rich internal structure.

Another unresolved corner case calls into question the general nature of the above ‘untensed’ heuristic, as at least the existential be construction can invoked in an untensed context, for example There fail to be bookcases in the office (csli/217). Abstractly, this example gives rise to a similar topology to the noun phrase coordination above: a non-scopal modifier on an instance variable ends up sharing its label not with predication introducing that variable (the cooordination conjunction or the bookcase predication, respectively), but rather with a predication that takes the instance as its (non-scopal) argument.

As a last disambiguation resort, EDS conversion might have to resort to surface order, i.e. prefer representative nodes that linearly precede their ‘competition’. Seeing as the relevant sets of modifiers in these constructions (in English) tend to be post-head, such a simple heuristic might actually come to the right result. However, it does not quite seem linguistically very well founded ...

EdsConversion (last edited 2016-07-12 23:56:29 by StephanOepen)

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