Diachronic proximity vs. data sparsity in cross-lingual parser projection. A case study on Germanic

For the study of historical language varieties, the sparsity of training data imposes immense problems on syntactic annotation and the development of NLP tools that automatize the process. In this paper, we explore strategies to compensate the lack of training data by including data from related varieties in a series of annotation projection experiments from English to four old Germanic languages: On dependency syntax projected from English to one or multiple language(s), we train a fragment-aware parser trained and apply it to the target language. For parser training, we consider small datasets from the target language as a baseline, and compare it with models trained on larger datasets from multiple varieties with different degrees of relatedness, thereby balancing sparsity and diachronic proximity. Our experiments show (a) that including related language data to training data in the target language can improve parsing performance, (b) that a parser trained on data from two related languages (and none from the target language) can reach a performance that is statistically not significantly worse than that of a parser trained on the projections to the target language, and (c) that both conclusions holds only among the three most closely related languages under consideration, but not necessarily the fourth. The experiments motivate the compilation of a larger parallel corpus of historical Germanic varieties as a basis for subsequent studies. 1 Background and motivation We describe an experiment on annotation projection (Yarowski and Ngai, 2001) between different Germanic languages, resp., their historical varieties, with the goal to assess to what extent sparsity of parallel data can be compensated by material from varieties related to the target variety, and studying the impact of diachronic proximity onto such applications. Statistical NLP of historical language data involves general issues typical for low-resource languages (the lack of annotated corpora, data sparsity, etc.), but also very specific challenges such as lack of standardized orthography, unsystematized punctuation, and a considerable degree of morphological variation. At the same time, historical languages can be viewed as variants of their modern descendants rather than entirely independent languages, a situation comparable to low-resource languages for which a diachronically related major language exists. Technologies for the cross-lingual adaptation of NLP tools or training of NLP tools on multiple dialects or language stages are thus of practical relevance to not only historical linguistics, but also to modern low-resource languages. The final paper will be published under a Creative Commons Attribution 4.0 International Licence (CC-BY), http: //creativecommons.org/licenses/by/4.0/.