Publication at Faculty of Mathematics and Physics |
2023
Abstract
As a result of the ongoing push for unification, extension and integration of morphological resources, need arises for reliable low-resource morph classification, especially root identification. The paper reports on our experiments with multiple root identification methods with various degrees of supervision, tested on several Indo-European languages, showing, among others, that given morphological segmentation, surprisingly good root identification can be achieved using simple unsupervised statistical methods, the main bottlenecks being compounding and homomorphy resolution.