Automatická morfologická disambiguace korpusů řady SYN: spolupráce lingvistické introspekce a strojového učení

The paper deals with the current method of automatic morphological tagging of corpora of contemporary Czech of the SYN series and other corpora within the Czech National Corpus. From SYN2020 onwards, the corpora are annotated on the basis of an improved concept. The paper starts with a brief description of newly introduced features concerning tokenization and lemmatization (introduction of sublemmata), and of the tagging of multiword tokens (i.e. compound forms like abys, cos); a new attribute, verbtag, is also presented. Then the successive steps of the entire annotation process are described. The core of the paper provides a detailed description of the procedure of automatic morphological disambiguation, namely the combination of two methodologically different approaches: the LanGr system of linguistically motivated disambiguation rules based on introspection, and the MorphoDiTa tool based on machine learning – we call this combination a hybrid approach. Particular emphasis is laid on the detailed characterization of the LanGr system, primarily on compiling specific lists of bigrams and trigrams of lemmas and forms labeled as global identifiers and on using these identifiers in disambiguation rules. The success rate of the hybrid system compared to the success rate of the stand-alone MorphoDiTa system is also presented and plans are briefly outlined for further development of our hybrid morphological tagging approach.