📄 Analysis of Corpus-based Word-Order Typological Methods

Summary

The authors investigate how different corpus-based methods of word-order typology group languages. They compare the resulting groupings with a classification based on a typological database. The material consists of the parallel PUD treebanks from UD v2.7 in CoNLL-U format for 20 languages. For each method, a language is represented as a vector of frequencies of particular word-order phenomena. Distances between languages are computed with the Euclidean metric, and the results are shown as clustering dendrograms. The baseline is the syntactic features from lang2vec. The authors compare three corpus-based approaches:

• the linear patterns of the Marsagram tool, i.e. the order of parts of speech within a single tree node;
• the relative position of head and dependent;
• the order of verb and object.

Each method yields its own grouping of languages. lang2vec and the head–dependent method come closest to the genealogical classification. The authors do not aim to identify the best approach. Instead, they show that each method highlights a different aspect of syntax. Corpus-based vectors make it possible to assess language similarity in finer detail than databases such as WALS. This is useful, among other things, for selecting related languages in NLP applications.

Key points

• Data: PUD contains 1,000 parallel sentences from news and Wikipedia for 20 languages, among them Czech, Polish and Russian. Most of the languages are Indo-European; the rest belong to 8 other families. The sentences were translated mostly from English (750), so "translationese" effects are possible.
• Procedure: frequency vectors, then Euclidean distances, dissimilarity matrices and hierarchical clustering. The results are compared with lang2vec, which draws on URIEL (including WALS).
• Marsagram: Icelandic turned out to be completely isolated, and Japanese also stands apart. Even Spanish and Portuguese do not join a common group. The positions of subject, verb and object do not play a decisive role here.
• Head–dependent: the Romance languages form a separate cluster. Icelandic falls in with the Slavic languages, which is consistent with type 10 in Hawkins (1983). Thai, Arabic and Indonesian form a separate subgroup.
• Verb–object: the languages split clearly into OV and VO. German, for which WALS records no dominant order, also ended up in the OV cluster.