Background

Elementary Dependency Structures (EDS; see the EdsTop page for general background) are designed as a lossy reduction of full MRSs, leaving out most of the underspecified scope information that, arguably, is among the more intricate parts of ERG semantics and that, arguably, makes a comparatively smaller contribution to ‘core meaning’. Thus, among the original EDS goals was simplification (removing MRS-specific ‘technicalities’), in order to relate to dependency-based approaches to meaning representation and to enable practical downstream tasks.

To give a few examples, full ERG MRSs require that propositional arguments are scopally sub-ordinated by a =q handle constraint; quantifiers always bind their restrictor via =q, leave their body unbound, and never share their label; intersective modifiers in attributive use (‘the fierce dog’) share the label of their head, but in predicative use (‘the dog is fierce’) must not do so. For an input semantics to succeed in generation, the above wellformedness constraints (and others) must be satisfied, and the ERG generator is not very robust to ‘deviant’ input structures; underspecification of the label topology is often not possible.

Generation from (formally less complex) EDS inputs, thus, would in all likelihood lower the burden for creators of semantics for ERG generation. For the above reasons, however, conversion from EDSs to full MRSs will require insertion of some additional (scope-related) information, beyond what is explicit in the EDS. It is an open question whether and to what degree this information can be automatically added to an input EDS or, in other words, how much of the scope underspecification mechanisms of the ERG can be predicted from basic predicate–argument dependencies and static information about the grammar (which may in principle include lexical information about individual predicates).

This page documents an ongoing experiment in this direction, aiming for a tiny target audience of ERG insiders in its present form.

General Approach

As of mid-November 2015, the LOGON infrastructure includes code to mechanically rewrite an EDS into an incomplete ‘pseud’-MRS, with one EP per node, logical variables introduced according to the ‘distinguished variable’ notion of Oepen & Lønning (2006), and one role per dependency edge. Assuming that the ERG SEM-I is available, format conversion can be invoked as follows:

  MRS(23): (lkb::do-parse-tty "she didn't sing.")
  MRS(24): (setf eds (ed-convert-edge (first *parse-record*)))
  {e10:
   x5:pron<0:3>{x PERS 3, NUM sg, GEND f, PT std}[]
   _1:pronoun_q<0:3>[BV x5]
   e10:neg<4:10>{e SF prop, TENSE untensed, MOOD indicative, PROG -, PERF -}[ARG1 e3]
   e3:_sing_v_1<11:16>{e SF prop, TENSE past, MOOD indicative, PROG -, PERF -}[ARG1 x5]
  }
  MRS(25): (eds-to-mrs eds)
  h1:e6:{ h9:"_sing_v_1_rel"<11:16>(x2 e8) h7:neg_rel<4:10>(e8 e6) 
          h5:pronoun_q_rel<0:3>(x2) h3:pron_rel<0:3>(x2) }
    { h1 qeq h7 }

It has long been true that variable properties from distinguished variables were attributed to corresponding EDS nodes, but the default textual serialization used to now show these; observe how the new output default in the LOGON code now is to show node properties, optionally prefixed with type a designator, enclosed between curly braces inbetween the node label, (optional) surface link, and the list of outgoing dependency edges (to revert EDS textual serialization to the old behavior, set *eds-show-properties-p* to nil).

In the MRS above, the argument of the negation is an eventuality, and the quantifier lacks its restrictor (which would prevent successful generation). To extend this structure into an ERG-compliant MRS (where the argument of the negation is scopal, for example), there is an emerging set of MRS rewrite rules, bundled as a stand-alone transfer grammar with the forthcoming 1214 release candidate of the ERG (see the sub-directory eds/ and, specifically, comments in the files mark.mtr and scope.mtr for the general approach). At present,

Interactive Experimentation

A convenient set-up for playful exploration of robust generation from EDSs is to (a) prepare a profile where the recorded semantics is reduced to EDS (rather than full MRS); (b) run one instance of the LOGON system as an ERG generator server; and (c) use another instance to retrieve EDSs via [incr tsdb()], apply the EDS transfer rules from the multi-MRS browser, and send transfer results to generation afterwards.

For step (a), the following command will generate a new profile mrs.d from the treebanked version of the MRS test suite included with the ERG:

  $LOGONROOT/redwoods --erg --thin --semantics eds --target mrs.d mrs

In the result relation of the new profile, the semantics field should now be populated with EDSs (note that the –semantics command line option is a recent addition to the redwoods script).

Step (b) follows the ‘classic’ recipe:

  (lmt)
  (rsa :erg t)

Similary, to launch the interactive development environment (step (c)), use the following:

  (lmt)
  (rsa :eds)

Once [incr tsdb()] has registered the mrs.d profile, use Browse | Results (in the second LOGON instance, i.e. the one loaded with the EDS transfer grammar) to obtain a summary of exemplars and associated syntactic and semantic analyses. Double-clicking on the red numbers in the MRS column will retrieve the recorded EDSs for that exemplar and invoke the multi-MRS browser on them; when given an EDS input structure, the browser will (now) automatically invoke the conversion to pseudo-MRS. Much like in interactive development for MT (see Step (5) on the LogonInstallation page), the Transfer button will invoke the EDS transfer grammar and display the result (Debug | Indexed will show a more readable view; Debug | Step supports step-wise inspection of intermediate transfer results). Finally, the Generate button will transmit the transfer result to the generator server, and with a bit of luck generation results will pop up (or helpful debugging messages appear in the ‘console’ window of the LOGON instance providing the generation server).

Known Limitations

The code for serialization of EDSs including node properties, for reading EDSs back in, for conversion from EDS to MRS, and for the new type of flow control in the EDS transfer grammar was added to the LOGON Tree in late November 2015 and will likely undergo further refinement; at present, generation from EDSs requires that the LOGON tree is run from source code.

To date, the EDS transfer grammar (probably) has only limited coverage of coordinate structures; is known to drop CARG values in intersective modifiers (e.g. cardinal adjectives); does not deal with sub-ordinated relative clauses (‘the dog that i think is fierce barked.’), the existential ‘be’ construction, and tag questions; and fails to resolve predicate names that (arguably illegitimately) appear as both types and strings in the ERG (e.g. _now_a_1). With these limitations (and doubtlessly others), nevertheless, generation is possible for a large proportion of exemplars from the MRS test suite; more systematic (batch) testing is pending.

Last update: 2015-11-25 by StephanOepen [edit]