A native DMRS grammar - i.e., the semantics is intended to be interpreted directly as a DMRS (without variables). *** Please keep this grammar in line with mrscomp-updated if you make any modifications to it! *** In order to try this grammar out, load into the LKB in the usual way via the script file. The DMRS option from the menu associated with the small parse trees in the LKB interface will show a DMRS structure extracted directly from the grammar, rather than via the usual conversion from MRS/RMRS. [ You may have to call (lui-shutdown) to get the LKB interface. Note that the mrsglobals file in this grammar sets (defparameter lkb::*dmrs-grammar-p* t) to redefine the behaviour of the DMRS menu item - you'll need to set it back if you are working with a normal grammar without restarting the LKB. ] This is based on the mrscomp grammar, but with some changes which are independent of making this a DMRS grammar: - added object control (persuade) and object raising (expect) - added an XARG to verbs to make control cleaner - changed from a DP analysis to an NP analysis - make infinitival `to' etc have no semantics rather than give them a dummy relation (actually, this is forced by the DMRS approach, but is anyway cleaner) Feb 15 2016 - all these changes are now in the mrscomp-updated grammar, so please see this for a detailed MRS/DMRS comparison DMRS encoding: The changes are all in the types.tdl and rules.tdl files (lexicon.tdl changes are due to the changes above). HCONS, LBL and all qeqs are gone. RELS is a dlist of nodes, monotonically accumulated as usual HOOK has INDEX, LTOP, XARG - all nodes node is a type with PRED and NODETYPE features (very similar to the old `relation' - not really necessary to create a different type but clearer) PRED is as mrscomp grammar (not the proper DMRS pred so look like gpred in XML) NODETYPE is ref-ind etc, as in mrscomp grammar - this could alternatively have been done by subtyping node but NODETYPE is probably useful for the finer-grained features such as tense etc which are usually associated with INDEX - FIX not done yet (the mrscomp grammar didn't have such features) node has subtypes introducing the features ARG1, etc and RSTR, which point to links link is a type with TARGET feature (which points to the to node) and subtypes corresponding to the post part of a DMRS link: eq-link neq-link h-link (could also have heq-link) RULES: a phrase always gets its INDEX from the HEAD - note that this is a stronger principle in a characteristic variable grammar than in previous approaches. It means, for instance, that the INDEX of `probably sleep' is the INDEX of `sleep', which is desirable. It means that the only exceptions to the generalisation that the INDEX must correspond to the lexical item which is the syntactic head of the phrase have to be cases where there is a null semantic item, such as a case-marking preposition, which transfers the INDEX (in favt the entire HOOK) from the non-head argument. This is crucial to understanding how DMRS behaves (relatively) similarly to dependency grammars, although the LTOP mechanism allows for the proper expression of constituency (e.g., the RSTR of quantifiers). See detailed notes. The treatment of LTOP depends on particular rule types. For head complement rules, the entire HOOK always comes from the head. For the specifier-head rule, the LTOP comes from the head - this is not so clearly correct for Det N combination - it relies on NEQ links (see detailed notes). Modifier rules have to be spilt accoding to whether the modifier is scopal (in which case it contributes the LTOP) or non-scopal, in which case the LTOP comes from the HEAD. Application of scopal/non-scopal is controlled by a subtype of pos. Nothing else about semantics happens in the rules in this grammar (there are no construction preds - FIX I probably should add one as an example). Note that this is, in some ways, more complex than in the old mrscomp grammar, where the entire HOOK was passed up to the head, but the need for this already arises in an MRS grammar using characteristic variables. In particular, since modifiers no longer share their INDEX with the modifiee, we can no longer regard the modifier as the (semantic) head. LEXICON: lexemes, for convenience, have a KEY as in the mrscomp grammar but this is not an appropriate feature for words and phrases. (As in the mrscomp grammar, there is a strict lexeme/word/phrase split with lexeme/word interconversion via inflectional rules or the const-pump rule.) - ordinary lexemes have a single node in the RELS corresponding to the KEY - this is also the INDEX. For non-quantifiers it is also the LTOP, quantifiers leave LTOP unbound (BUT - see detailed notes). - no-sem-lexems have nothing on the RELS list and no INDEX etc. Note that this is a necessary difference from a non-DMRS grammar where they might have an INDEX which is later shared. It implies that a node without semantics that contributes to properties such as tense has to do this by accessing the INDEX of a syntactic argument. - all NP arguments are linked with neq-links which access the INDEX - all scopal arguments are h-links (COMPs, SPEC or MOD) which take an LTOP (assumption - an NP is never a scoped over argument - see detailed notes) - all intersective mod arguments access the INDEX - except in the case of an NP they are eq-links auxiliary verbs are treated as simply scopal - this is to be FIXed additional stuff for control and raining cases - the link between the SPR's INDEX for a normal verb and its XARG is set up in a lexical type subj-verb - srv verbs do not link their SPR to one of their own ARGS but share their SPR with the COMP's SPR (as usual) - this gives the raising effect without any explicit coindexation of semantic arguments other than the standard scopal argument - orv verbs are similar, but a COMP is shared with another COMP's SPR - scv and ocv verbs link the XARG of a COMP to one of their ARGs in addition to the usual linking TO DO: - agreement and tense as semantic properties, including changing treatment of auxiliaries and making code in dmrscomp extract features (as MRS) - simple conjunction - predicative adjectives - note that these have to have an NEQ link with the noun phrase. - adjective modification - it seems etc - simple example of a construction contributing semantics, maybe compounds - quantifier modification (maybe ...) - relative clauses including the cases where we currently have undirected EQ links - make generation work ... Note that INSTLOC is absent - it should be unnecessary for this grammar. - figure out what this means for ERG!