'If you've heard it, you can say it': towards an account of expressibility
David D. McDonald, Charles F. Greenbacker · 2010
We have begun a project to automatically create the lexico-syntactic resources for a microplanner as a side-effect of running a domain-specific language understanding system. The resources are parameterized synchronous TAG Derivation Trees. Since the KB is assembled from the information in the texts that these resources are abstracted from, it will decompose along those same lines when used for generation. As all possible ways of expressing each concept are pre-organized into general patterns known to be linguistically-valid (they were observed in natural text), we obtain an architectural account for expressibility. 1. Expressibility People speak grammatically. They may stutter, restart, or make the occasional speech error, but all in all they are faithful to the grammar of the language dialects they use. One of the ways that a language generation system can account for this is through the use of grammar that defines all of the possible lexico-syntactic elements from which a text can be constructed and defines all their rules of composition, such as lexicalized Tree Adjoining Grammar (TAG). Without the ability to even formulate an ungrammatical text, such a generator provides an account for human grammaticality based on its architecture rather than its programmer. We propose a similar kind of accounting for the problem of expressibility: one based on architecture rather than accident. Expressibility, as defined by Meteer (1992), is an issue for microplanners as they decide on which lexical and syntactic resources to employ. Not all of the options they might want to use are available in the language – they are not expressible. Consider the examples in Figure 1, adapted from Meteer