Linguistically possible translation

[51] introduces the notion linguistically possible translation which provides a useful methodological concept for work in Machine Translation (see also [53]). A system which implements the notion `linguistically possible translation' employs linguistic knowledge only. Clearly, in order to produce the correct or best translation such a system has to be augmented with other (artificial intelligence) components. However, at the current state of technology it seems unrealistic to expect that these other components can be constructed in the near future. Therefore, a more realistic goal for MT consists of the construction of systems implementing the notion `linguistically possible translation'. Such systems may also be of practical interest, because it may be possible to augment such restricted systems with a component that interacts with the user -- for example in the case of difficult disambiguation problems. Thus the user can provide help in order to extract the best translation from the computed set of linguistically possible translations. It is assumed that the overall best translation is in fact to be found (in a significant number of cases) among this set. This chapter therefore will be focused on the notion `linguistically possible translation', rather than on translation in general.

Consider a class of MT systems which employ linguistic knowledge only. In such a system, a source text is assigned a set of meaning representations according to the rules of the grammar of the source language. On the basis of such a meaning representation the target grammar then produces a set of target sentences for this grammar. Such a system thus produces a set of linguistically possible translations. [51] assumes that the relation `linguistically possible translation' (lpt) is a symmetric relation. Thus, t<sub>target</sub> lpt t<sub>source</sub> iff t<sub>source</sub> lpt t<sub>target</sub>. On the other hand, the relation `best of linguistically possible translations' is not. For example assume that disambiguation based on knowledge of the world is outside the linguistic components. The asymmetry occurs if the translation of some (unambiguous) source sentence is ambiguous, and where furthermore the `added' reading of the target sentence is to be preferred for extra-linguistic reasons. An example can be constructed using the famous Bar-Hillel sentence, cf. figure 5.1.

Figure 5.1: Best translation is not symmetric. In the English to Dutch direction, extra-linguistic knowledge dictates the second sentence.

This view of translation is extremely poor. For example, it does not take world knowledge into account as we saw above. Moreover, there are many other factors that could be taken into account in defining linguistically possible translations, e.g. preservation of style, (indirect) speech act, honorifics, etc. It is hoped (and expected) that an approach based on the poor view described can be useful as a basis for future richer views.

An important question of translation is whether there always is a meaning-preserving translation. It may be the case that there are meanings in one language that are not expressible at all in some other language (for some discussion cf. [42,47]). It may even be the case that one cannot know whether the meaning expressed in two languages is the same (cf. [70]). These are important questions, but the approach outlined here does not depend on how they are answered. Our approach is concerned only with the case where the same meaning can be expressed in both languages. Our question is `how to describe possible translations', not `is translation possible'.

The symmetry of the linguistically possible translation relation provides the motivation for reversible MT systems [51,44]. If the lpt relation from language l₁ to l₂ is in fact the same relation as going from l₂ to l₁, then it seems very natural to try to characterize this relation only once -- and to construct a program which is able to compute this relation, given this single characterization, in both directions.

Such an approach has a number of advantages, most of which coincide with the advantages already mentioned in chapter 1. Thus, a reversible architecture has

• theoretical advantages. Given that there is only one linguistically possible translation relation between two given languages, then we should describe this relation only once.
• practical advantages. If we are interested in building systems that translate between some given languages then it may be expected that writing a single component for a language pair may in fact be `cheaper' than writing two separate components.
• methodological advantages. It can be argued that the goal of writing a reversible translation system improves the quality, even in one direction. This point is worked out by [35], for the translation between English and French.