Other issues

A detailed version of guidelines can hardly be produced before the syntax group has discussed the status of syntax information, which has to be decided for an architecture proposal, i.e.

• The main entities of the model
• The articulation between the main entities

The architecture group will need to continue its activity as an infrastructural service for the other groups over the whole period of the EAGLES exercise, to ensure that these recommendations remain consistent with those of the other task groups, and to amend them if necessary.

There should be more work, in both the morphosyntax group and the architecture group, on problems of the representation of multiword units; the following two reasons need to be kept track of:

• even though CGPMU may provide a straight forward way of representing certain types of multiword units, it is not sure that this device is sufficient to cater for the wide variety of phenomena likely to occur. It would be useful to set up a list of the facts about multiword items which need to be treated.
• EAGLES should not commit too early to a given descriptive device for multiword units: it may be better to keep the guidelines flexible and to be able to insert, later, some proposal which may be based on more and also more detailed analysis of phenomena from different languages.

Hierarchies, Classification, Inheritance in the lexicon

The architecture group considers the introduction of classifcations and, along with these of inheritance hierarchies a useful structuring device on top of flat (and maybe redundant) dictionaries. However, it is less evident whether EAGLES should ``prescribe'' the use of this device or even of a given instance of it.

Proposal:

• EAGLES recommendations ideally come in two layers:
  • a flat layer of basic lexical descriptions, which may be somewhat redundant, because it does not make use of any classificatory devices;
  • a structuring layer, which introduces classifications, generalizations, etc., and which may involve certain representational ways of expressing these generalizations, e.g. through inheritance hierarchies.
• If an EAGLES user introduces classifcations at the second level, this should be done via specification of relations or generalization rules which express a two-way correspondence between the flat representations and the hierarchical ones. See the discussion about morphological dictionary (base level) and corpus analysis tagsets (applied, second level) in the minutes of the morphology group.

It is assumed that

• different descriptive levels may make use of classificatory devices with different properties (e.g. monotonic vs. non- monotonic inheritance hierarchies);
• different applications of EAGLES lexicons may later introduce generalizations of their own.

If we think in terms of larger actual dictionaries, it would evidently be useful to have some generalization and classification already at the level of basic lexical descriptions, e.g. to avoid redundancy in the different word forms of a lemma (have lemma-specific information instead, and additional ideosyncratic information on word forms separately). The question there is which classifications may usefully be proposed by EAGLES: evidently, there is no single best generalization over a heterogenous set of objects, which is why a generalization process on a ``data-only'' basis is not possible. Usually, generalizations at this level are based on the classificatory assumptions of the approach advocated.