Experiments in learning models for functional chunking of Chinese text

This paper introduces a system of chunk-like annotation to describe Chinese predicate-argument structures, and describes some of our work in developing learned models for automatically annotating fresh text according to this system. The annotation is very similar in form to other chunking systems, except that chunks are defined not bottom-up but top-down, in terms of relationship to a main predicate. Bottom-up parsing of these structures seems to require great consideration of structural information and long-distance influences. Explicit representation of chunk structure during parsing allows us to provide more informative features, and experiments show that these give significant improvements in performance