INFDok - Dokumentenvolltextserver des Fachbereichs Informatik

Kurzfassung auf Englisch:

The problem of (semi-)automatic treebank conversion arises when converting between different schemas, such as from a language specific schema to Universal Dependencies, or when converting from one Universal Dependencies version to the next. We propose a formalism based on top-down tree transducers to convert dependency trees. Building on a well-defined mechanism yields a robust transformation system with clear semantics for rules and which guarantees that every transformation step results in a well formed tree, in contrast to previously proposed solutions. The rules only depend on the local context of the node to convert and rely on the dependency labels as well as the PoS tags. To exemplify the efficiency of our approach, we created a rule set based on only 45 manually transformed sentences from the Hamburg Dependency Treebank. These rules can already transform annotations with both coverage and precision of more than 90%.