1 LFG qua concurrent constraint programming

By now, LFG is a well-established framework for natural language analysis based on the interrelated notions of structure, description and correspon-dence(Kaplan, 1995; Kaplan and Bresnan, 1981). The key idea is to analyze an utterance simultaneously along several dimensions — constituent structure, functional structure, semantic structure, with provision for additional dimensions. The analyses on each dimension are not independent, rather they are tied together via constraints — pieces of partial information, about some finite aspects of interest. For instance, the consistuent-structure (c-structure) of the utterance is usually described by means of a (somewhat liberal) context-free grammar. Simultaneously, via the parsing process, constraints are assembled that specify a projection from the c-structure into a functional structure (f-structure) that describes the grammatical relations between different constituents (e.g., subj, obj), using attribute-value matrices. As (Dalrymple et al., 1993) shows, it is possible to augment such an architecture by introducing a semantic projection (projection) from the f-structure. Interrelationships between semantic projections are specified by means of formulas in a " glue logic " (taken to be linear logic). Conceptually, deductions with these formulas are performed simultaneously with the elaboration of the c-structure and the f-structure, resulting, on a successful analysis, with the simultaneous synthesis of the c-, f-and-structures of the original utterance. A very attractive and appealing aspect of LFG' s architecture is the separation of concerns. Different processing techniques and different vocabularies of constraints are appropriate for different aspects of analysis. It is not necessary to dictate the primacy of one dimension of analysis over another, or to use tech-2 Chapter 1. LFG qua concurrent constraint programming niques from one dimension to simulate the techniques of another. Nevertheless it is possible for the analysis along these dimensions to be interrelated since each produces constraints that can influence others. This notion of constraint-based communication at the heart of its architecture is what crucially supports LFG' s elegant compositionality, and makes it a prime exemplar of a general paradigm of compositional computing (Saraswat, 1997). Constraints enjoy several remarkable properties (?). First, they specify partial information — a constraint (e.g., X + Y Z) need not uniquely specify the value of its variables , only impose certain restrictions on them. Second, they are additive: the order of imposition of constraints does not matter. Third, they are rarely inde-accumulating more constraints can yield much more information than processing them separately. Fourth, they are non-directional: typically a constraint on …