Multi-component tree adjoining grammars, dependency graph models, and linguistic analyses

Abstract

Recent work identifies two properties that appear particularly relevant to the characterization of graph-based dependency models of syntactic structure 1 : the absence of interleaving substructures (well-nestedness) and a bound on a type of discontinuity (gap-degree ≤ 1) successfully describe more than 99% of the structures in two dependency treebanks (Kuhlmann and Nivre 2006). 2 Bodirsky et al. (2005) establish that every dependency structure with these two properties can be recast as a lexicalized Tree Adjoining Grammar (LTAG) derivation and vice versa. However, multicomponent extensions of TAG (MC-TAG), argued to be necessary on linguistic grounds, induce dependency structures that do not conform to these two properties (Kuhlmann and Möhl 2006). In this paper, we observe that several types of MC-TAG as used for linguistic analysis are more restrictive than the formal system is in principle. In particular, tree-local MC-TAG, tree-local MC-TAG with flexible composi-1 Whereas weak equivalence of grammar classes is only concerned with string sets and fails to shed light on equivalence at the structural level, our work involves the equivalence of derivations and graph based models of dependencies. Thus, our work is relevant to certain aspects of grammar engineering that weak equivalence does not speak to. 2 These properties hold for many of the so-called nonprojective dependency structures and the corresponding noncontext free structures associated with TAG, further allowing CKY type dynamic programming approaches to parsing to these dependency graphs. tion (Kallmeyer and Joshi 2003), and special cases of set-local TAG as used to describe certain linguistic phenomena satisfy the well-nested and gap degree ≤ 1 criteria. We also observe that gap degree can distinguish between prohibited and allowed wh-extractions in English, and report some preliminary work comparing the predictions of the graph approach and the MC-TAG approach to scrambling.