next up previous contents
Next: Constraint-based transfer Up: The architecture of MiMo2 Previous: The architecture of MiMo2

Other constraint-based approaches to MT

The objective to build a reversible MT system using a series of reversible unification grammars is reminiscent of the CRITTER system [36], the TFS [112], the CLE [2] and ELU [76]. In CRITTER logic grammars are being used; Emele and Zajac use a type system including an inheritance mechanism to define transfer-like rules. In CLE and ELU the semantic representations defined by monolingual grammars are put in correspondence by special transfer rules, rather than using one formalism for both monolingual and bilingual grammars (as we propose). In section 5.6 I describe an approach to certain non-compositional translation cases which seems not available to the two latter formalisms, but which is available in the MiMo2 architecture.

A somewhat different approach is advocated in [41]. In that approach a system is described where an LFG grammar for some source language is augmented with equations that define (part of) the target level representations. A generator derives from this partial feature structure a phonological representation according to some LFG grammar of the target language. Instead of a series of three grammars this architecture thus assumes two grammars, one of which defines both the source language and the relation with the target language. The translation relation is not only defined between semantic representations but may relate all levels of representation (c-structure, f-structure, $ \sigma$-structure). Although in this approach monolingual grammars may be used in a bidirectional way it is unclear whether the translation equations can be used bi-directionally. Furthermore, the approach faces problems with certain non-compositional translations as already discussed by the authors and in some more detail in [78].

[111] discusses a constraint-based approach to MT in which translation relations are exclusively defined by relating lexical entries of the source and target language. A translation in this approach proceeds, somewhat simplified, as follows. A source text is built from a `bag' (i.e. a multi-set) of source language lexical entries. Each of these lexical entries may be related to lexical entries of the target language, giving rise to a bag of target lexical entries. The target text is computed by generating (non-deterministically) a sentence using this bag of lexical entries (and the combination rules of the target grammar). In relating the lexical entries of different languages, several pieces of information may be put in correspondence; most notably semantic information. The target sentences that are generated use this semantic information. This approach works thanks to a crucial property of the grammars, inherited from UCG [113]: all semantic representations are projected from the lexicon. However, the approach faces severe difficulties for grammars in which semantic information is not always projected from lexical entries. Furthermore, if this approach to translation is generalized in order for other types of information (such as `style') to be put in correspondence as well, then we need to assume that such information is also projected from lexical entries; it seems problematic to assume that this is always possible or linguistically satisfactory.


next up previous contents
Next: Constraint-based transfer Up: The architecture of MiMo2 Previous: The architecture of MiMo2
Noord G.J.M. van
1998-09-30