How Does the Granularity of an Annotation Scheme Influence Dependency Parsing Performance

Abstract

The common use of a single de facto standard annotation scheme for dependency treebank creation leaves the question open to what extent the performance of an application trained on a treebank depends on this annotation scheme and whether a linguistically richer scheme would imply a decrease of the performance of the application. We investigate the effect of the variation of the number of grammatical relations in a tagset on the performance of dependency parsers. In order to obtain several levels of granularity of the annotation, we design a hierarchical annotation scheme exclusively based on syntactic criteria. The richest annotation contains 60 relations. The more coarse-grained annotations are derived from the richest. As a result, all annotations and thus also the performance of a parser trained on different annotations remain comparable. We carried out experiments with four state-of-the-art dependency parsers. The results support the claim that annotating with more fine-grained syntactic relations does not necessarily imply a significant loss of accuracy. We also show the limits of this approach by giving details on the fine-grained relations that do have a negative impact on the performance of the parsers.