Yet Another Format of Universal Dependencies for Korean

Proceedings of the 29th International Conference on Computational Linguistics, pages 5432–5437 October 12–17, 2022. 5432 Yet Another Format of Universal Dependencies for Korean Yige Chen†∗ Eunkyul Leah Jo‡∗ Yundong Yao‡∗ KyungTae Lim¶ Miikka Silfverberg‡ Francis M. Tyers§ Jungyeul Park‡ †The Chinese University of Hong Kong, Hong Kong ‡The University of British Columbia, Canada ¶Hanbat National University & TeddySum, South Korea §Indiana University, USA †yigechen@link.cuhk.edu.hk ‡{eunkyul,iameleve}@student.ubc.ca ¶ktlim@hanbat.ac.kr §ftyers@iu.edu ‡{msilfver,jungyeul}@mail.ubc.ca Abstract In this study, we propose a morpheme-based scheme for Korean dependency parsing and adopt the proposed scheme to Universal De- pendencies. We present the linguistic rationale that illustrates the motivation and the necessity of adopting the morpheme-based format, and develop scripts that convert between the orig- inal format used by Universal Dependencies and the proposed morpheme-based format au- tomatically. The effectiveness of the proposed format for Korean dependency parsing is then testified by both statistical and neural models, including UDPipe and Stanza, with our care- fully constructed morpheme-based word em- bedding for Korean. MORPHUD outperforms parsing results for all Korean UD treebanks, and we also present detailed error analyses. 1 Introduction Dependency parsing is one of the tasks in nat- ural language processing that have been investi- gated extensively. Using the dependency gram- mar, it finds the relations between the words in a sentence, and forms an acyclic dependency graph that explains the grammatical structure of the sen- tence. With various machine learning techniques, the current dependency parsers are able to ap- proach human performances given English cor- pora.1 Previous studies trying to probe or improve Korean dependency parsers are lacking, especially for those considering the underlying reason and rationale based on the linguistic properties of Ko- rean. The parsers do not perform as well as their English counterparts, partly due to the fact that Korean is a language that has more complicated linguistic features that make parsing on the word level difficult. ∗Yige Chen, Eunkyul Leah Jo, and Yundong Yao con- tributed equally. ¶Corresponding author. 1https://ai.googleblog.com/2016/05/ announcing-syntaxnet-worlds-most.html There have been previous studies trying to cope with the word-level representation issues of Ko- rean (Choi and Palmer, 2011; Park et al., 2013; Kanayama et al., 2014). Given that Korean is an agglutinative language that heavily relies on morphemes, and the natural segmentation does not correctly reflect either the words or the mor- phemes of Korean texts, Park and Tyers (2019) suggested an annotation scheme that decomposes Korean texts into the morpheme level, and applied the morpheme-based format to POS tagging. In this study, we propose a morpheme-based scheme for Korean dependency parsing that is developed based on Park and Tyers (2019), and adopt the proposed scheme to Universal Depen- dencies (Nivre et al., 2016, 2020), which con- tains two Korean dependency parsing treebanks, namely the GSD treebank (McDonald et al., 2013) and the Kaist treebank (Choi et al., 1994; Chun et al., 2018). While the two Korean treebanks meet the standards of Universal Dependencies and have been studied for dependency parsing tasks exten- sively (Kondratyuk and Straka, 2019; Qi et al., …

Content truncated to 500 words