A cross-genre morphological tagging and lemmatization of the Russian poetry: distinctive test sets and evaluation
The poetic texts pose a challenge to full morphological tagging and lemmatization since the authors seek to extend the vocabulary, employ morphologically and semantically deficient forms, go beyond standard syntactic templates, use non-projective constructions and non-standard word order, among other techniques of the creative language game. In this paper we evaluate a number of probabilistic taggers based on decision trees, CRF and neural network algorithms as well as a state-of-the-art dictionary-based tagger. The taggers were trained on prosaic texts and tested on three poetic samples of different complexity. Firstly, we suggest a method to compile the gold standard datasets for the Russian poetry. Secondly, we focus on the taggers’ performance in the identification of the part of speech tags and lemmas. We reveal what kind of POS classes, paradigm classes and syntactic patterns mostly affect the quality of processing.