This grammar is derived from the mrscomp grammar and is essentially
identical to it, except as necessary to produce (a variant of) RMRS directly.

1. decomposed relations

main relations are type relation, argument relations are type arg-relation.
relation and arg-relation are subtypes of ep.  PRED value of
an arg-relation is always a subtype of argsort: these are types.

The first element on the RELS list for a lexical entry is taken to be 
the main relation and is coindexed with KEY, for convenience in setting up
the lexicon.

Lexical types are defined to have a list of relations, all with the same LBL

trans-verb := verb-lxm &
[ SPR < [HEAD.FORM norm, SEM.HOOK.INDEX #arg1, SEM.HOOK.LTOP #ing] >,
  COMPS < phrase & [HEAD det, HEAD.FORM norm, SPR <>, COMPS <>, 
                    SEM.HOOK.INDEX #arg2, SEM.HOOK.LTOP #ing] >,
  SEM.HOOK.LTOP #ing,
  SEM.RELS <! relation & [ LBL #h], [ LBL #h, PRED arg1, VAR #arg1 ], 
                        [ LBL #h, PRED arg2, VAR #arg2 ] !> ].


2. INGs

ING are added for accumulation of ING constraints.
ING append just like HCONS and RELS.  

The approach to ING adopted in RASP-RMRS is to create INGs between
modifier and modifiee.  The MRS->RMRS conversion code simulates
this. It would be possible to do this directly in these grammars too,
but would mean considerable modification - we would no longer be able
to unify LTOPs.  Experiment here with an alternative which gives a
slightly different notion of INGs (should be interconvertable).

ING constraints relate handles - CNJ is the label of the (implicit)
conjunction, CNJED is the label of the relations being conjoined.

e.g., every big bad dog - l0:big(x), l1:bad(x), l2:dog(x), ing(l3,l0),
ing(l3,l1), ing(l3,l2), l4:every(x), l4:rstr(h5), l4:body(h6), h5 qeq l3

equivalent to: l0:big(x), l1:bad(x), l2:dog(x), l3:conj(l0,l1,l2),
l4:every(x), l4:rstr(h5), l4:body(h6), h5 qeq l3

e.g., bark loudly - l1:bark(e), l1:ARG1(x), l2:loud(e), ing(l3,l1), ing(l3,l2)

The LTOP of non-quantifiers is in an ING with the LBL of the main (KEY) 
relation

non-quant-lexeme := lexeme & 
[ SEM.HOOK [ INDEX #index,
             LTOP #ltop ],
  SEM.ING <! [ CNJ #ltop, CNJED #lbl ] !>,
  KEY [ LBL #lbl, ARG0 #index ]].

This is the only way that INGs are added in the grammar.

The effect of this is to have vacuous INGs - effectively this is as though
everything is projecting a conjunction and in many cases it will be 
unnecessary

e.g., every dog - l2:dog(x), ing(l3,l2), l4:every(x), l4:rstr(h5), l4:body(h6), h5 qeq l3

equivalent to: l2:dog(x), l3:conj(l2), l4:every(x), l4:rstr(h5), l4:body(h6), h5 qeq l3

The experimental code in mrsoutput converts INGs to explicit n-ary
conjunctions and removes vacuous ones.

This is a more serious manipulation than we usually do in conversion
to an MRS from the grammar FSs and I am not exactly happy about this.

What are the implications for generation?

The structures with conj explicit will scope, although to test this I
have converted the quantifiers back to MRS style since otherwise
considerable modification would be needed to the code that looks for
free variables etc.

e.g., l0:big(x), l1:bad(x), l2:dog(x), l3:conj(l0,l1,l2),
l4:every(x, h5, h6), h5 qeq l3