Topological Field Parsing for Discourse Modeling in German
By Jackie C.K. Cheung, University of Toronto
Abstract:
Topological fields are a model of clausal syntax for freer-word-order languages such as German, in which grammatical role is not the sole determinant of word order. Existing work on modeling coherence in text relies on grammatical role distinctions which correlate well with information status distinctions in fixed-word-order languages like English, but less so in freer-word-order ones like German. In this talk, we argue that topological fields are a better indicator of information status than grammatical role in German. We first show that topological fields can be parsed at high accuracies using existing statistical parsing technology, and then show that they can be fruitfully integrated into a model of local coherence, outperforming both grammatical role and simple word order information in a sentence ordering experiment. Applying the model to a natural language generation setting, we also integrate topological field information into a constituent order generation system and improve generation performance over the baseline system.