Invited talk: Dependency Representations, Grammars, Folded Structures, among Other Things!

of invited talk Aravind K. Joshi Department of Computer and Information Science and Institute for Research in Cognitive Science University of Pennsylvania Philadelphia PA USA [email protected] In a dependency grammar (DG) dependency rep­ resentations (trees) directly express the depen­ dency relations between words. The hierarchical structure emerges out of the representation. There are no labels other than the words them­ selves. In a phrase structure type of representa­ tion words are associated with some category la­ bels and then the dependencies between the words emerge indirectly in terms of the phrase structure, the non­terminal labels, and possibly some indices associated with the labels. Behind the scene there is a phrase structure grammar (PSG) that builds the hierarchical structure. In a categorical type of grammar (CG), words are as­ sociated with labels that encode the combinatory potential of each word. Then the hierarchical structure (tree structure) emerges out of a set of operations such as application, function composi­ tion, type raising, among others. In a tree­adjoin­ ing grammar (TAG), each word is associated with an elementary tree that encodes both the hi­ erarchical and the dependency structure associ­ ated with the lexical anchor and the tree(s) asso­ ciated with a word. The elementary trees are then composed with the operations of substitution and adjoining. In a way, the dependency potential of a word is localized within the elementary tree (trees) associated with a word. Already TAG and TAG like grammars are able to represent dependencies that go beyond those that can be represented by context­free grammars, but in a controlled way. With this perspective and with the availability of larger dependency annotated corpora (e.g. the Prague Dependency Treebank) one is able to as­ sess how far one can cover the dependencies that actually appear in the corpora. This approach has the potential of carrying out an ‘empirical’ inves­ tigation of the power of representations and the associated grammars. Here by ‘empirical’ we do not mean ‘statistical or distributional’ but rather in the sense of covering as much as possible the actual data in annotated corpora! If time permits, I will talk about how dependen­ cies are represented in nature. For example, grammars have been used to describe the folded structure of RNA biomolecules. The folded structure here describes the dependencies be­ tween the amino acids as they appear in an RNA biomolecule. One can then ask the question: Can we represent a sentence structure as a folded structure, where the fold captures both the depen­ dencies and the structure, without any additional labels? * Part of this work is in cooperation with Joan Chen Main, University of Pennsylvania, Philadelphia PA and Johns Hopkins University Baltimore, MD.