An Ontology-based Approach to Automatic Part-of-Speech Tagging Using Heterogeneously Annotated Corpora

In the LOD era, the conceptual interoperability of language resources is established by using modular architectures like the Ontologies of Linguistic Annotations (Chiarcos, 2008a, OLiA). Available as a part of the Linguistic Linked Open Data (LLOD) cloud,1 OLiA provides ontological representations of annotation schemes for over 70 languages, as well as their linking to a reference model. We successfully train an ontology-based POS tagger on corpora with different tag sets of divergent granularity and partially compatible annotations. Making use of OLiA, we achieve interoperability of annotation schemes, and, despite sparse training data, we do not only outperform state-of-the-art POS taggers in concept coverage, but also show how traing on heterogeneously annotated data produces richer morphosyntactic annotation with no or only marginal loss of precision.