-
Notifications
You must be signed in to change notification settings - Fork 4
LogonVpm
In the MRSs derived by grammars like the ERG, a fair amount of information is encoded as variable (or index) properties, i.e. attributes and values associated to semantic variables. TENSE, MOOD, and NUM(ber) are typical such properties, and often they facilitate an encoding of information that remains relatively close to morpho-syntactic, grammar-internal distinctions. For the external interface to each grammar, it seems desirable to aim for a unified, yet simple inventory of properties and values, shared across grammars (to the extent of actually being relevant for a token language). At the same time, a grammar may have internal constraints on the organization of morpho-syntactic information, and often there may be distinctions made in a grammar that need not be projected into the external, semantic interface. The Variable Property Mapping facility is a bi-directional tool that relates grammar-internal and -external encoding on the basis of a declarative specification of the mapping relations.
The types of variables themselves often require adjustment in creating an MRS from a feature structure (AVM), or vice versa. In the ERG, for example, the grammar-internal type names are event and ref-ind, while the canonical MRS variable types are e and x. Traditionally, such correspondences in the LKB were established by means of the built-in function determine-variable-type(), which in turn builds on a family of globals (named, not co-incidentally, similar to the ERG-internal choices, e.g. *event-type* and *ref-ind-type* for the above example). This original mapping of variable types in the LKB was uni-directional, i.e. it is applied when constructing an MRS from a feature structure, but not in the inverse direction (that means, in principle, that information cast as variable types is not utilized in generation until the very end, i.e. the post-generation semantic compatibility check).
As of early 2009, variable type correspondences can be encoded as part of the VPM too. The ERG, for example, now includes in its semi.vpm:
event <> e
ref-ind <> x
individual <> i
handle <> h
non_event <> p
* >> u
semarg << u
This section has to precede all variable property mappings (see below) in the file. To make VPM-based variable type mapping replace determine-variable-type(), it is necessary to further set a new MRS global (typically in the file mrsglobals.lsp):
(setf *variable-type-mapping* :semi)
With the variable type correspondences as part of the VPM, the grammar no longer needs to provide MRS globals *event-type* et al. Note, however, that (just like in the original LKB machinery) VPM-based variable type mapping is at present only applied in forward direction (see comments in vpm.lsp, generate.lsp, and transfer.lsp for the remaining issues that require LKB backwards compatibility in this respect).
Consider a grammar which records information about person and number in a single sortal hierarchy, with values like 1sg, 2per, et al. The grammar makes use of a feature PN as the home for the combined person and number values, and for grammar-internal reasons, PN is embedded below a feature PNG inside of the feature structures that correspond to semantic variables in the MRS (keep in mind that MRSs are not feature structures, yet for each grammar there is a known bijection between the AVM and the MRS universes). Following is a mapping to convert the grammar-internal, conflated encoding of person and number into an external representation that is compatible with the MRS best practice in LOGON and the RMRS DTD.
PNG.PN : PERS NUM
1sg <> 1 sg
1pl <> 1 pl
1per <> 1 !
1per << 1 *
2sg <> 2 sg
2pl <> 2 pl
2per <> 2 !
2per << 2 *
3sg <> 3 sg
3pl <> 3 pl
3per <> 3 !
3per << 3 *
* >> ! !
! << * *
The above example defines a set of rules that map one (or more) properties into one or more properties. When reading an MRS off the AVM produced by the grammar, for each correspondence of (groups of) properties, values are compared to sub-rules in order, until the first match: at that point, output values are inserted into the result set of properties. Processing of subsequent rules continues against the original properties, so that there could be multiple matches: the PNG.PN to PERS and NUM decomposition, thus, could also be done in two separate rule sets. At the end of the day, however, only properties resulting from successful matches will be included in the output MRS, i.e. everything not explicitly carried over will be suppressed.
VPMs can be applied in two directions: forward application, mapping from the left-hand side of rules into the right-hand side, and backward application, producing the inverse mapping. While the prototypical mapping rule is bi-directional, there may be special cases. The VPM machinery provides the following operators:
-
<> defines an equivalence mapping, i.e. operates in either direction
>> provides a forward-only mapping, i.e. in left to right direction
<< provides a backward-only mapping, i.e. in right to left direction
By default, values are compared by subsumption (against the type hierarchy of the grammar, for the time being). There are variants of the mapping operators that use equality for testing instead. These are ==, =>, and <=, respectively.
Furthermore, there are a few special operators for matching and outputting of values. In a nutshell, * will match any (existing) value in the input or insert whatever was matched into the output; conversely, ! will not insert the corresponding property into the output, e.g. the sub-rule from our example above
1per <> 1 !
will have the effect of only inserting [PERS 1] into the output, while the NUM property will be omitted. When applied in backward direction, ! can be used to match absence of the corresponding property, but this usage is really only useful in conjunction with additional properties: without conditioning on another property, a rule matching on the absense of a value would insert the output property on variables of all types. A potentially more useful matching operator that facilitates the insertion of default values conditions a rule on the type of the embedding variable. The ERG, for example, includes the following mapping of TENSE values:
E.TENSE : TENSE
past <> past
present <> pres
future <> fut
real_tense <> tensed
untensed <> untensed
* >> untensed
untensed << *
untensed << [e]
Here the effect of the last two sub-rules, which are limited to the backward direction, is to (a) convert all values not matched by earlier rules into untensed and (b) insert untensed on all variables of type e that have no TENSE property already. In other words, the [e] match operator subsumes the use of ! but additionally conditions on a specific variable type, so as to avoid inserting TENSE values into referential indices, say.
A grammar can include any number of VPMs; the function mt:read-vpm() can be used in the LKB `script' file to load a VPM and associate an identifier to it, e.g.
(mt:read-vpm (lkb-pathname (parent-directory) "semi.vpm") :semi)
The distinguished identifier :semi will activate its VPM for the read-out of MRSs from parsing results (where the VPM is applied in forward direction) and, conversely, for the reverse mapping of MRSs given as input to the generator. Additional VPMs can be put to use in transfer, as an input or output filter for example.
In PET, there is provisional support for loading the SEM-I VPM, i.e. the VPM that applies to the extraction of MRSs from feature structures obtained from parsing results; add a statement like the following to your settings file:
vpm := "semi".
To provide multi-VPM support in PET, this interface will likely change in the future.
The duality of grammar-internal and external (aka SEM-I) naming conventions creates a number of challenges for processing that, conceptually, operates on MRSs but at the same time needs to make use of the type hierarchy, e.g. to test for compatibility or subsumption (aka underspecification) relations of MRS predicates, variable types, and variable properties. The ultimate solution will require that the SEM-I provies complete information about MRS-level relations that the grammar wants to make available in its interface (in principle, the SEM-I could also provide additional layers of hierarchical relations, i.e. ones that hold at the MRS level but are not part of the grammar-internal hierarchy). The SEM-I could then be imported into its own hierarchy, possible in a private namespace per grammar, such that multiple MRS-level hierarchies can be loaded in parallel, e.g. corresponding to source and target language in transfer.
But this vision is not yet implemented in the LKB MRS core and transfer machinery, and hence the current technology (as of early 2009) takes some 'short-cuts' in the creation and application of transfer rules, as well as in MRS-related parts of the generator. In debugging, it may be necessary to understand in detail the duality of MRS naming conventions, and the following paragraph provides a brief summary of the current state of play.
For transfer rules (and thus trigger rules in generation), we require the internal hierarchy of property values. Therefore, the :semi VPM is not applied during the creation of transfer rules, except for its variable type correspondences (see above), if any. Unlike variable properties, the application of transfer rules to an MRS assumes canonical (external) variable types, e.g. h, e, x, et al. This is primarily a matter of convenience: grammars tend to have large sets of intermediate variable types, and in testing an input MRS which only provides information in the much simpler, external hierarchy, the extra grammar-internal type distinctions would at best be a nuisance. Accordingly, grammars need to provide, as part of their type hierarchy, the canonical MRS variable types too, i.e.
u := *top*.
i := u.
p := u.
h := p.
e := i.
x := i & p.
The situation is parallel for generation. To instantiate and Skolemize lexical entries, the input MRS is mapped through the VPM in reverse direction, except for variable types. In the current setup (and the LKB generator ever since its inception), variable types are not considered during lexical instantiation. Thus, in principle, semantic distinctions cast in the input semantics by virtue of variable types are not considered by the generator until very late in the process, viz. the post-generation semantic compatibility test. For all we know, however, existing grammars hardly, if at all, make use of variable types in this manner; for example, a grammar would have to provide distinct lexical entries, both with the same predicate but incompatible variable types on their ARG0, say.
There is a discussion of how to delete and attribute for only some predicates at NoJa.
Home | Forum | Discussions | Events