Using Constraint Grammar for Chunking

This paper presents and evaluates a novel and flexible chunking method using Constraint Grammar (CG) rules to introduce chunk edges in corpus annotation. Our method exploits preexisting (non-constituent) morphosyntactic annotation such as part-of-speech or function tags, but can also be made to work on raw text, integrated with other CG modules. The first version of the chunker was developed for German CG-annotated interview data, with a parallel English version derived from the German one, indicating a high degree of language-independence of the rules in the presence of generalized syntactic-functional tags (e.g. subject, object, modifier). Two different approaches are discussed, one for minimal, flat chunking, the other for deep, nested chunking. The system has a reasonable performance and robustness for both, achieving F-scores of 89.1 and 97.4 for nested and minimal chunking, respectively. Xml markup is supported, and with a full set of rules, the tool can be used to convert CG annotation into complete constituent trees in VISL or TIGER format.