Tagging Complex Non-Verbal German Chunks with Conditional Random Fields

Abstract

We report on chunk tagging methods for German that recognize complex non-verbal phrases using structural chunk tags with Conditional Random Fields (CRFs). This state-of-the-art method for sequence classification achieves 93.5% accuracy on newspaper text. For the same task, a classical trigram tagger approach based on Hidden Markov Models reaches a baseline of 88.1%. CRFs allow for a clean and principled integration of linguistic knowledge such as part-of-speech tags, morphological constraints and lemmas. The structural chunk tags encode phrase structures up to a depth of 3 syntactic nodes. They include complex prenominal and postnominal modifiers that occur frequently in German noun phrases.