A unification grammar defined in grammar formalisms such as DCG [27] and PATR [37]
usually defines a relation between a string of words and some representation, sometimes
called logical form. In sign-based approaches such as UCG [47] and HPSG
[28], the string of words is not assigned a privileged status but is represented
as the value of one of the attributes of a feature structure.
In this approach a unification
grammar defines a set of feature structures where each feature structure represents a
pairing of a logical form and a string.
It is also possible to use unification grammars to define other relations between
feature structures. In MiMo2 unification grammars are used
to encode bilingual knowledge too: each (bilingual)
unification grammar defines a symmetric transfer relation.
Monolingual unification grammars
define relations between strings and logical forms; bilingual grammars
define relations between (language specific) logical forms.
In MiMo2 the translation relation between two natural languages is defined by a series of three unification grammars. Each of these grammars is reversible; the system consisting of three of such grammars applied in
series is reversible too [44].
The monolingual components of the formalism thus consist of
unification grammars, similar to PATR.
Unlike PATR the terminal elements in the formalism are
not defined in the lexicon, but orthographical, inflectional and
morphological rules define the relation between the terminals and
a lexicon of stems and affixes. For example, the word `eaters'
is analyzed into [eat,er,s] by the orthographical component. For
the orthographical component we use a reversible two-level system
[32][5][21].
Reversible inflectional rules relate to a list of stems and
affixes
with the feature structure
.
Inflection is defined by a formalism comparable to the
paradigmatic approach of [7].
Morphological
analysis is based on a separate reversible unification grammar in which derivational
processes and compounding can be defined [3]. For example,
could be analysed as
. Note
that the separation of inflectional rules and compound/derivation
rules implements a type of `level' theory defended by e.g. [1].
The unification grammars defined by the user in a PATR like style are compiled into
Prolog (an extension to the compilation described in [14]),
to enable an efficient implementation of parsing, generation and transfer.
The parser is a
`left-corner' parser augmented with a well formed substring table
and a reachability table [25][26].
Both techniques are optimized by using a set of
`restricted' features [34]. The generator of Mimo2 is a
member of the generation family described in [39][43][38][42].
Transfer is implemented as a top-down backtrack search procedure.