The data-oriented parsing approach: Theory and application

Abstract

Parsing models have many applications in AI, ranging from natural language processing (NLP) and computational music analysis to logic programming and computational learning. Broadly conceived, a parsing model seeks to uncover the underlying structure of an input, that is, the various ways in which elements of the input combine to form phrases or constituents and how those phrases recursively combine to form a tree structure for the whole input. During the last fifteen years, a major shift has taken place from rule-based, deterministic parsing to corpus-based, probabilistic parsing. A quick glance over the NLP literature from the last ten years, for example, indicates that virtually all natural language parsing systems are currently probabilistic. The same development can be observed in (stochastic) logic programming and (statistical) relational learning. This trend towards probabilistic parsing is not surprising: the increasing availability of very large collections of text, music, images and the like allow for inducing statistically motivated parsing systems from actual data. A corpus-based parsing approach that has been quite successful in various fields of AI, is known as Data-Oriented Parsing or DOP. DOP was originally developed as an NLP technique but has been generalized to music analysis, problem-solving and unsupervised structure learning [7, 8, 14, 81]. The distinctive feature of the DOP approach, when it was first presented, was to model sentence structures on the basis of previously observed frequencies of sentencestructure fragments, without imposing any constraints on the size of these fragments. Fragments include, for instance, subtrees of depth 1 (corresponding to context-free rules), as well as entire trees. © 2008 Springer-Verlag Berlin Heidelberg.