Introduction: Goals of the workshop

A number of key researchers in the field were nominally invited to make a presentation or at least participate in thematic discussions. The number of participants was kept low primarily to foster fruitful exchanges and to promote discussion that could lead to concrete conclusions. In addition, budgetary considerations prevented offering travel subsidies, which also contributed both to the small scale of the event and to its being linked to another event at which most of the people we wanted to include would already be present.

The goals of the workshop were to provide key researchers in the field an opportunity for brain-storming on the definition of the concepts involved in the question of the linquistic adequacy of formalisms for natural language processing. We expected each participant both to draw on their experience to provide information on the way some of these concepts are already realized in actual grammar formalisms and to bring in their own opinions on the way these concepts should be realized in future formalisms.

Thus we asked participants to address the following theoretical key topics:

• How do we determine linguistic adequacy?

  Adequacy of the formalism is not the same as coverage of a grammar written in that formalism. How can we distinguish between adequacy with respect to the language data and adequacy with respect to a linguistic theory?

• From the point of view of the coverage of the grammars: are the existing formalisms for NLP expressive enough?

  Although the problem of testing is not central to our theme, since there is an EAGLES Working Group on ``Evaluation and Assessment'', we also have to adress the question of measuring adequacy.

• Are the formalisms committed to a particular theory? Up to what degree does the formalism of your choice allow you to express generalizations?

  For instance, levels for thematic roles (why don't some formalisms use them more, although they are expressible?); word order (how natural is it to treat a free word-order language?) ; binding principles (do the same principles allow one to treat different phenomena, or does one need separate accounts for them?).

• Taking the position that formalisms are indeed powerful enough, how can we reduce complexity in the face of problems which increase complexity? Can we process large grammars with such complex formalisms?

• From a processing point of view: How can weights be used? Are weights obtainable by training? Can we induce grammars from input?

• On the question of degradation of data: How do existing formalisms handle unforeseen input? What counts as ill-formed? Are there several levels of ill-formedness?

• Questions of modularity, scalability and reusability, both in theory and in practice.

Each of these topics can be articulated on two levels, as questions of competence vs. performance: adequacy at the level of formalisms and complexity of formalisms vs. the level of specific applications and adequacy from the point of view of processing.

• List of Participants