@inProceedings{klironomou-etal-2026-exploring-367379,
title = {Exploring the syntactic parsing of learner texts in Modern Greek: Variety trumps size, genre trumps grammaticality},
abstract = {We investigate which properties of train data affect the performance of syntactic parsers on learner texts. We train and evaluate six models for Modern Greek, using the same annotation scheme (Universal Dependencies) and the same parser (MaChAmp), but varying train and test sets. The data come from standard texts, learner texts and learner with errors corrected. We confirm that training and testing on the same language variety yields better results than, for instance, training on standard texts and testing on learner texts. Moreover, our experiment setup makes it possible to compare the role of variety to that of other factors and dissect the very notion variety into more specific factors. Most importantly, we show that training on the same variety has stronger effect than train set size, and that adding non-target variety to the train set does not necessarily decrease performance. In practical terms, it means that adding learner data (that are often scarce) to the standard train set is likely to boost the performance on learner data without damaging the performance on standard data. Further, we show that the same value of grammaticality across train and test does not have an effect, while having the same genre does.},
booktitle = {Proceedings of the 15th Workshop on Natural Language Processing for Computer Assisted Language Learning},
author = {Klironomou, Christina and Masciolini, Arianna and Berdicevskis, Aleksandrs},
year = {2026},
publisher = {University of Tartu Library},
pages = {130--139},
}