What Treebanks Can Do For You: Rule-based and Machine-learning Approaches to Anaphora Resolution in German

This paper compares two approaches to computational anaphora resolution for German: (i) an adaption of the rule-based RAP algorithm that was originally developed for English by Lappin and Leass, and (ii) a hybrid system for anaphora resolution that combines a rule-based pre-filtering component with a memory-based resolution module. The data source is provided by the TüBa-D/Z treebank of German newspaper text (Telljohann et al., 2003) that is annotated with anaphoric relations. This treebank uses as its data source a collection of articles of the German daily newspaper taz (die tageszeitung). Due to their fine grained syntactic annotation, the TüBa-D/Z treebank data are ideally suited as a basis for the identification of markables for pronominal reference and for extracting relevant syntactic and semantic properties for each markable. The TüBa-D/Z annotation scheme distinguishes four levels of syntactic constituency: the lexical level, the phrasal level, the level of topological fields, and the clausal level. The primary ordering principle of a clause is the inventory of topological fields, which characterize the word order regularities among different clause types of German and which are widely accepted among descriptive linguists of German (cf. e.g. Höhle (1986)). The TüBa-D/Z annotation relies on a context-free backbone (i.e. proper trees without crossing branches) of phrase structure combined with edge labels that specify the grammatical function of the phrase in question. Figure 1 shows an example tree from the TüBa-D/Z treebank for sentence (1). The sentence is divided into two clauses (SIMPX), and each clause is subdivided into topological fields. The main clause is made up of the following fields