The multilingual architecture

The LU in the lingusitic architecture is seen as an operational notion. It identifies meanings within a language. In other words, the LUs of a particular language support and organize the semantics of this language. Although the LUs and their organizations are different in each language, the LU can be considered as an anchoring for multilingual relations. Equivalences between languages are obtained by drawing relations between their specific LUs according to relational rules explained hereafter.

Equivalence between LUs of different languages is multi-bilingual, transfer-based and unidirectional (See Annex number 3 ``THE MULTILINGUAL ARCHITECTURE SCHEMA").

Multi-bilingual

Each dictionary in a lexical resource complying with the MULTILEX standard is basically monolingual. From each LU of each monolingual dictionary, it is possible to specify equivalences to LUs of several other monolingual dictionaries.
Example:

DE "Fahrrad-lu..." EN "bicycle-lu..." NL "fiets-lu..." FR "bicyclette-lu..."

Each bilingual equivalence is dealt with separately. The features describing an equivalence for a given LU to an LU of another language is collected in a separate block and each block is attached to this LU. If several equivalents of the same LU exist in the other language, then more than one feature blocks can be created to account for each equivalence relation.
Example:

DE "Fahrrad-lu..." EN ``bicycle-lu...'' EN "bike-lu..." EN "cycle-lu..." NL "fiets-lu..." NL "rijwiel-lu..." FR "bicyclette-lu..." FR "vélo-lu..."

Conversely, when two LUs in language L1 has the same equivalent in language L2, then both the LUs of L1 point to the same LU of L2.
Example:

Transfer-based

The equivalence system is based on a transfer model. Each monolingual dictionary deals with the description of LUs according to monolingual criteria, and transfer provides the means to accommodate the source information to the target language needs. The notions of "source" and "target" dictionary are only valid from the point of view of transfer. Otherwise, each monolingual dictionary can be source and target. The bilingual equivalence block contains - along with the target LU - bilingual contrastive information, i.e., meaning discrepancies between the source and target LUs and conditions for syntactic matching. It also contain a characterization of the equivalence relation which can be full, partial in several ways or nil for different reasons.

Although LUs are semantic objects, the transfer is mainly syntax-based. This is not a restriction imposed by the standard but it results from the fact that large lexical resources do not provide extensive semantic descriptions yet which could supply enough information to perform semantic-based matching. The syntactic part of the standard provides the basic features from which to derive an interface format for bilingual syntactic matchings including elementary semantic features to apply selection restrictions.

Unidirectional

For each bilingual transfer the source LU is considered from the point of view of the target language and transfer blocks are made up according to the target dictionary LUs. Note, however, that the source language description at the semantic level always remains unaltered. The fact that new transfer entries or target languages are added, never affects the source LU.
The testing material - used for disambiguating the meaning of the source LU against the target language - and the directions used for the accommodation of the source syntactic structure to the target syntactic structure apply to one language pair and one direction only.

The consequence is that the resulting transfer information is not reversible. Such an organization has been proposed, mainly for reasons of feasibility. There is also an organizational advantage to this structure: each language community, participating in a project involving multilingual data can have the organizational and managerial responsibility over one source language and one unidirectional language pair.

However, it should be clearly stated that the architecture deals with the linguistic and logical organization of the data, and that it leaves out problems relevant to applications. For example, the transfer blocks can be automatically reversed for a number of LUs marked as fully equivalent when specific translation lexicons are extracted from the main dictionary.