Large Scale Semantic Construction for Tree Adjoining Grammars

Developing a Tree Adjoining Grammar which contains the information necessary for building the basic compositional semantics of sentences is a highly complex engineering task. To ensure consistency, ease of writing, of maintainance and of debugging, it is therefore important that this information be described at the appropriate level of abstraction. In the first part of this paper (sections 2 and 3), we show how to achieve a highly factorised integration of semantic information into a Tree Adjoining Grammar (TAG, [1]) using a particularly expressive grammar formalism recently developed by [2]. The second part of the paper shows how the resulting TAG can be used to support semantic construction that is, to associate the sentences generated by the grammar with a semantic representation. Contrary to other linguistic frameworks such as Lexical Functional Grammar, Head Driven Phrase Structure Grammar or Categorial Grammar, there is no clear consensus in TAG on how to perform semantic construction. This is because Tree Adjoining Grammar associates a derivation not with one, but with two structures namely, a derivation tree and a derived tree; and because it is unclear which of these two structures best supports semantic construction. As TAG elementary trees localise predicateargument dependencies so that derivation trees resemble semantic dependency trees, the TAG derivation tree has long been taken to provide an appropriate basis for semantic construction. Nevertheless, it has repeatedly been shown that the derivation tree alone does not provide all the information needed to perform