Coarse-grained vs. fine-grained Lithuanian dependency parsing

Advances in Intelligent Systems and Computing, vol. 1224, ISSN:2194-5357. Scopus 2019-SNIPIn this paper we present the dependency parsing experiments done on the Lithuanian treebank Alksnis v2.1. This treebank contains 84 different fine-grained syntactic labels (and can be generalized to 10 coarse-grained), is annotated with lemmas, parts-of-speech and other fine-grained morphological information (cases, numbers, tenses, etc.). During our experiments with the transition-based MaltParser system we have explored an impact of different morphological features. The best results were achieved when the following features were taken into account: case, verb form, named entity, word type, mood on the coarse-grained dependency labels and case, verb form, aspect, named entity and person on the fine-grained. The increase in LAS (UAS) over the baseline was ∼ 7.16% ( ∼ 4.77%) and ∼ 8.03% ( ∼ 5.19%) on the coarse-grained and the fine-grained reaching ∼ 71.90% ( ∼ 75.45%) and ∼ 71.14% ( ∼ 76.20%), respectivelyKauno technologijos universitetasTaikomosios informatikos katedraVytauto Didžiojo universiteta