Reversible Transfer of Context-sensitive Translations

The purpose of this section is to show that the formalism proposed for transfer is more powerful than some previously defined transfer formalisms such as the CAT framework [6,102], but also some of the constraint-based transfer formalisms [2,76], even taking into account the constraint on transfer grammars defined in the previous section. Furthermore, this extra power is required to handle some non-compositional translations. As in monolingual uses of constraint-based grammars we may percolate all kinds of information in the feature structures, for example to define context-sensitive translations.

In formalisms such as CAT, a transfer rule essentially translates a tree by translating parts of the tree recursively. The rules thus always operate on parts of the input object; this input object cannot be modified. This leads to complex rules for the treatment of context-sensitive translations. For example, the English adjective `strong' is normally translated into `sterk' in Dutch. In the case of `strong criticism', however, the translation has to be `scherpe kritiek' (sharp criticism). Assuming that this regularity has to be treated in transfer, the only possible way to obtain this result in CAT is to define a transfer rule that translates the structure in which both `strong' and `criticism' occur. This is problematic in cases where this larger structure contains other parts that have to be translated as well; the translation of these other parts can however also be irregular. This leads to special rules to handle the combination of such irregular cases (see also [7]). It may also be problematic if the representation of the adjective and the noun do not appear as sisters in the representation but may be arbitrarily far away from each other (e.g. because other adjectives intervene). In that case a rule has to be written for each different possibility. In cases where there is no limit to the distance even this escape will not work.

The current framework allows a compositional treatment of such context-sensitive cases because we simply can percolate information using the constraints. The validity of this claim is established with the following an example. First, recall that the argument structures for noun phrases such as `very strong whisky', look as follows:

$$\avm{ \mbox{\it sort}: \mbox{\rm modifier}\\ \mbox{\it mod}... ...red}: \mbox{\rm whisky}\\ \mbox{\it num}:\mbox{\rm sg} } }$$
The basic rule to translate argument structures of sort `modifier' is defined as in the following rule (this rule replaces the preceding rule for modifier structures). In this rule the information of what the `head' of a modified structure is, is percolated through the features `gbhead' and `nlhead'. This value will then be equated with the features `gbheaded' and `nlheaded' that are associated with the modifiers.

The rule translating nouns such as `whisky' is defined as:

Now we are ready to define a special rule for the translation of `strong' into `scherp' if the adjective is headed by the noun `criticism'/`kritiek'.

Hence, the parse tree of the translation of `very strong unmotivated criticism' can be given as in figure 5.7.

Figure 5.7: Translating `very strong unmotivated criticism'

We thus showed how the use of `extra' constraints allows for a compositional (in fact reversible) treatment of context-sensitive translations.