Probabilistic induction of metrical trees for word stress assignment

An algorithm for word stress assignment in German compounds is introduced. First, a metrical tree is automatically derived from adjacent morpheme cohesion scores which are based on co-occurrence statistics. This tree is used to identify the stressed compound part by applying the compound stress rule of metrical phonology. Then the stressed syllable is identified within this compound part by means of a k-nearest-neighbor classifier using weighted vowel quantity and syllable coda type features. The accuracy of the metrical compound analysis amounted to 83%. Compound stress assignment was successful in 95%, and stress location within multi-syllable word stems in 84% of all cases