Relational-Realizational Syntax: An Architecture for Specifying and Learning Morphosyntactic Descriptions

This paper presents a novel architecture for specifying rich morphosyntactic representations and learning the associated grammars from annotated data. The key idea underlying the architecture is the application of the traditional notion of a "paradigm" to the syntactic domain. N-place predicates associated with paradigm cells are viewed as relational networks that are realized recursively by combining and ordering cells from other paradigms. The complete morphosyntactic representation of a sentence is then viewed as a nested integrated structure interleaving function and form by means of realization rules. This architecture, called Relational-Realizational, has a simple instantiation as a generative probabilistic model of which parameters can be statistically learned from treebank data. An application of this model to Hebrew allows for accurate description of word-order and argument marking patterns familiar from Semitic traditional grammars. The associated treebank grammar can be used for statistical parsing and is shown to improve state-of-the-art parsing results for Hebrew. The availability of a simple, formal, robust, implementable and statistically interpretable working model opens new horizons in computational linguistics --- at least in principle, we should now be able to quantify typological trends which have so far been stated informally or only tacitly reflected in corpus statistics.