Semantic layer

The semantic layer of the model deals with two distinct levels of representation: a first level may be seen as strictly lexical semantics, whereas the second one aims at representing a more cognitive type of contents.

The main entities for lexical semantics are SEMANTIC UNITS (USEM). They are closely connected with the syntactical layer, as every USEM has to be related to at least one USYN. This relation may be restricted through constraints about the USYN itself and/or by filters on the positions it governs in a construction ('productive' collocations can be described by such filters) ; the relation may also precise the way semantic arguments of a predicate match syntactic positions governed by an USYN, and give default semantic values for implicit arguments (for example when an absolute construction is possible).

An USEM may be connected to a linguistic PREDICATE, which can synthetize semantic information about predicative USEMs. This connexion may take various modalities: a PREDICATE may be lexicalized by one or more USEM, so may its arguments. For example, USEM1 cans be one of the lexicalizations of PREDICATE P, USEM2 may be the typical lexicalization of argument i for P, etc.

USEMs can be described:

• analytically through a set of values of semantic features that range from 'classical' componential feature to pragmatic feature referring to a scenario, and include mentions of domain, connotative value, etc.
• by means of cross-references that can be expressed through a set of specialized relations:
  • paradigmatic relations as hypo/hyperonymy, synonymy, meronymy, etc.
  • 'semantic derivation', that may represent derivation according to the meaning (with or without morphologic motivation), or information related to 'typicallity' (e.g. USEM1 is the typical location for USEM2 activity, USEM3 is the typical instrument for USEM4 action, etc.), when not expressed through a PREDICATE
  • relations of collocation preference (e.g. USEM1 is the preferred intensifier for USEM2, USEM3 is the support verb for such aspect of USEM4 action, etc.)

PREDICATEs may also be described in the two manners, although the features and relations available for them are less varied.

The semantic layer should also offer the possibility to abstract cognitive units (CONCEPTs) from USEMs and/or from PREDICATEs. Such units may be useful to factorize information about a lexical class of equivalence (for instance synonyms carrying different connotative values), but also to materialize units of content that do not have a lexical actualization for a given language: such 'lexical gaps' may need to be filled for the purpose of establishing taxonomies of for representing terminological data.

Finally, although there is not within GENELEX for the time being a complete specification for a multilingual level of description, the structure of the semantic layer paves the way for:

• a contrastive description of transfer information from lexical semantics in one language to lexical semantic in another language (using 'filters' in the same manner as when relating USYNs to USEMs)
• an interlingua description based upon sharing of CONCEPTs, even if this prospect seems rather fuzzy by now for direct NLP applications, as such a possibility could turn out to be very useful for terminological purposes.