Abstract
We investigate to what extent inflection can be automatically separated from derivation, just based on the word forms. We expect pairs of inflected forms of the same lemma to be closer to each other than pairs of inflected forms of two different lemmas (still derived from a same root, though), given a proper distance measure.
We estimate distances of word forms using edit distance, which represents character-based similarity, and word embedding similarity, which serves as a proxy to meaning similarity. Specifically, we explore Levenshtein and Jaro-Winkler edit distances, and cosine similarity of FastText word embeddings.
We evaluate the separability of inflection and derivation on a sample from DeriNet, a database of word formation relations in Czech. We investigate the word distance measures directly, as well as embedded in a clustering setup.
Best results are achieved by using a combination of Jaro-Winkler edit distance and word embedding cosine similarity, outperforming each of the individual measu