Weighing the Utility of Types in Grammatical Description

While computational linguists developing broad-coverage grammars share many assumptions about the formal devices that are necessary or useful for grammatical description, there is a distinct lack of agreement about the utility of types as first-class objects within such a grammar. In this talk I identify some of the motivations for employing types when implementing a large grammar, and illustrate the uses of these abstractions in the various tasks of the grammarian during the construction and testing of this complex resource. The primary motivations are grounded both in linguistics and in engineering: the linguistic concerns include the familiar ones of capturing generalizations (both monolingual and multilingual), increasing breadth of coverage of phenomena, maintaining declarativity for both analysis and generation, striving for parsimony in descriptive devices, and including compositional semantics as part of the grammar; the engineering issues include enabling successive refinement in grammar construction, balancing precision and robustness, reducing obstacles to efficient processing, and defining an interface for applications. The particular tasks that engage the grammar writer, and for which I argue that the use of types offers demonstrable benefit, include defining lexical selection and valence alternations, defining phrase structure rules and principles, linking syntax and semantics, refining and extending analyses of phenomena, identifying intended analyses (using statistical approximation), and correcting errors in definitions during the long process of grammar construction. I conclude with a consideration of some of the alternatives to the use of types in large grammar implementation, and suggest that these alternatives all fall short of providing the theoretical and practical benefits which typed feature structure grammars can enjoy.