Abstract
We present pilot experiments on splitting and identifying Czech compound words. We created an algorithm measuring the linguistic similarity of two words based on finding the shortest path through a matrix of mutual estimated correspondences between two phonemic strings.
Additionally, a neural compound-splitting tool (Czech Compound Splitter) was implemented by using the Marian Neural Machine Translator framework, which was trained on a data set containing 1,164 hand-annotated compounds and about 280,000 synthetically created compounds. In compound splitting, the first solution achieved an accuracy of 28% and the second solution achieved 54% on a separate validation data set.
In compound identification, the Czech Compound Splitter achieved an accuracy of 91%.