Evaluating Automatically Acquired F-structures against Propbank

f-structure representations is presented by Burke et al. (2004b). The annotation algorithm is the basis for the automatic acquisition of wide-coverage and robust probabilistic approximations of LFG grammars (Cahill et al., 2004) and for the induction of subcategorisation frames (O'Donovan et al., 2004; O'Donovan et al., 2005). Annotation quality is, therefore, extremely important and to date has been measured against the DCU 105 and the PARC 700 Dependency Bank (King et al., 2003). The annotation algorithm achieves f-scores of 96.73% for complete f-structures and 94.28% for preds-only f-structures against the DCU 105 and 87.07% against the PARC 700 using the feature set of Kaplan et al. (2004). Burke et al. (2004a) provides detailed analysis of these results. This paper presents an evaluation of the annotation algorithm against PropBank (Kingsbury and Palmer, 2002). PropBank identifies the semantic arguments of each predicate in the Penn-II treebank and annotates their semantic roles. As PropBank was developed independently of any grammar formalism it provides a platform for making more meaningful comparisons between parsing technologies than was previously possible. PropBank also allows a much larger scale evaluation than the smaller DCU 105 and PARC 700 gold standards. In order to perform the evaluation, first, we automatically converted the PropBank annotations into a dependency format. Second, we developed conversion software to produce PropBank-style semantic annotations in dependency format from the f-structures automatically acquired by the annotation algorithm from Penn-II. The evaluation was performed using the evaluation software of Crouch et al. (2002) and Rie-zler et al. (2002). Using the Penn-II Wall Street Journal Section 24 as the development set, currently we achieve an f-score of 76.58% against PropBank for the Section 23 test set.