Towards the Conversion of National Corpus of Polish to Universal Dependencies

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 5308–5315 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 5308 Towards the Conversion of National Corpus of Polish to Universal Dependencies Alina Wróblewska Institute of Computer Science, Polish Academy of Sciences Warsaw, Poland alina@ipipan.waw.pl Abstract The research presented in this paper aims at enriching the manually morphosyntactically annotated part of National Corpus of Polish (NKJP1M) with a syntactic layer, i.e. dependency trees of sentences, and at converting both dependency trees and morphosyntactic annotations of particular tokens to Universal Dependencies. The dependency layer is built using a semi-automatic annotation procedure. The sentences from NKJP1M are first parsed with a dependency parser trained on Polish Dependency Bank, i.e. the largest bank of Polish dependency trees. The predicted dependency trees and the morphosyntactic annotations of tokens are then automatically converted into UD dependency graphs. NKJP1M sentences are an essential part of Polish Dependency Bank, we thus replace some automatically predicted dependency trees with their manually annotated equivalents. The final dependency treebank consists of 86K trees (including 15K gold-standard trees). A natural language pre-processing model trained on the enlarged set of (possibly noisy) dependency trees outperforms a model trained on a smaller set of the gold-standard trees in predicting part-of-speech tags, morphological features, lemmata, and labelled dependency trees. Keywords: Polish Dependency Bank, National Corpus of Polish, Universal Dependencies, conversion 1. Introduction Natural language pre-processing (i.e. tagging, morphologi- cal analysis, lemmatisation, syntactic parsing, etc.) is a cru- cial issue in natural language understanding (NLU). It is an open question whether it is possible to infer syntactic and semantic knowledge essential for NLU without any su- pervision. If effective language-independent unsupervised algorithms are developed, the quality of pre-processing will undoubtedly improve and it will also be possible to pre- process less resourced languages. However, the state-of- the-art natural language processing (NLP) tools are still data-driven and a lot of attention and financial support is devoted to building high-quality training and testing data. Manual annotation of resources that can be used for training NLP models is expensive and time-consuming. Therefore, alternative ways of obtaining training data are being sought. One of them is self-labelling technique, e.g. Triguero et al. (2015), in which a model trained on a gold-standard data set (mostly a small set of training examples) is used to predict pseudo-training instances. A new model is then trained on both gold-standard data and automatically predicted data, assuming that the model will provide more accurate predic- tions, because it is trained on a larger set of (possibly noisy) training instances. We use the idea of self-labelling to build a large collec- tion of Polish dependency trees. A set of Polish sentences is automatically parsed with a dependency parser trained on the largest bank of Polish gold-standard trees. A new parsing model is then estimated on both the gold-standard trees and the automatically predicted ones. In order to re- duce the number of errors, the dependency trees are pre- dicted …

Content truncated to 500 words