📄 Comparative Concepts or Descriptive Categories: a UD Case study

Summary

The authors examine whether it is justified in UD to label the cases of different languages with the same names. They build on the distinction proposed by Haspelmath: a descriptive category belongs to a particular language, whereas a comparative concept is needed for typology. UD in effect uses the same label in both roles. Each case in a treebank is represented as a syntactic profile, i.e., the normalized distribution of the dependency relations by which a word in that case attaches to its governor. The profile is computed in two ways: over all parts of speech or over NOUN only. The data are taken from UD 2.14. First, the authors compare the cases of two closely related languages, Czech and Russian, using a nearest-neighbor graph. Next, they test whether identically named cases from all treebanks form distinct clusters, applying t-SNE, ToMATo and k-NN. Finally, they build the same kind of profiles for prepositions. Across languages, only the frequent cases are more or less recognizable, while the rest overlap heavily. For cross-lingual transfer this is an important caveat: an identical feature name does not guarantee identical behavior.

Key points

• In UD, the Case feature is used in at least three different ways. Mostly it is assigned to inflected nominal forms. It is also assigned to prepositions, and in Czech CLTT the value is simply copied from the noun. In Persian Seraji, only Loc, Tem and Voc occur: they mark adverbs of place and time and an interjection.
• Czech and Russian cases were compared by Euclidean distance in a 1-NN graph. When profiles are computed over all words, only the nominatives and genitives turned out to be mutual neighbors. For noun profiles, the accusatives and instrumentals were added to them. The datives, locatives and the Russian partitive remained confused.
• In the k-NN confusion matrix (k = 11) for Acc, Gen, Loc, Nom, 227 nominatives, 156 genitives and 130 accusatives were correctly recognized. Only 29 of 155 locatives were recognized, so the rarer cases get lost among the frequent ones.
• Neither t-SNE nor the persistence-based clustering ToMATo produced clearly separated clusters, even for the Nom/Gen pair.
• The profiles of French prepositions were averaged over all French corpora. Dans, par, sur, sous and vers behave similarly to the locative, avec to the instrumental or comitative. In pour, à and sans, a noticeable share of advcl appears due to infinitival constructions.