Automatic Morpheme Segmentation and Labeling in Universal Dependencies resources

摘要

Newer incarnations of the Universal Dependencies (UD) resources feature rich morphological annotation on the word-token level as regards tense, mood, aspect, case, gender, and other grammatical information. This information, however, is not aligned to any part of the word forms in the data. In this work, we present an algorithm for inferring this latent alignment between morphosyntactic labels and substrings of word forms. We evaluate the method on three languages where we have manually labeled part of the Universal Dependencies data-Finnish, Swedish, and Spanish-and show that the method is robust enough to use for automatic discovery, segmentation, and labeling of allo-morphs in the data sets. The model allows us to provide a more detailed morphosyntactic labeling and segmentation of the UD data.