This paper describes several compilation techniques that can transform an arbitrary feature structure-based grammar into one which is more efficient for parsing and generation. These techniques increase the performance of unification by introducing new dependent disjunctions into the grammar, pushing the disjunctions down as far as possible and then minimizing the interactions between disjunctions. Any unification algorithm that can handle disjunctive or dependent feature structures should benefit from these techniques. In fact, the technique for minimizing interactions between disjunctions is very general and can even be applied to Prolog programs. 1 Introduction There are two facts that conspire to make the treatment of disjunction an important consideration when building a natural language processing (NLP) system. The first fact is that natural languages are full of ambiguities, and in a grammar many of these ambiguities are described by disjunctions. 1 The second fact is that th...

The desirability of Horn clauses in logical deductive systems has long been recognized. The reasons are at least threefold. Firstly, while inference algorithms for full logics of any reasonable extent are typically intractable, for systems restricted to Horn clauses the picture is much better. (For example, in ordinary propositional logic, while the full satisfiability problem is NP-complete, a linear-time algorithm exists for Horn clauses.) Secondly, the knowledge-representation capabilities of Horn clauses, while weaker than those of the full logic, remain remarkably rich; indeed, far richer than that of simple conjunctive logic alone. Thirdly, Horn clauses define the maximal subset of a full logic which has the property of admitting generic models, which roughly means that for any set of Horn clauses, there is a least model of the clauses in that set. It is the purpose of this paper to initiate an investigation of Horn clause logic for an extended class of feature structures. After laying the groundwork for this context, we provide two key results. In the first, we show how the property of admitting