Data Augmentation for Low-Resource Languages in Multilingual Dependency Parsing
The result's identifiers
Result code in IS VaVaI
<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11320%2F26%3AZALZ4Y9G" target="_blank" >RIV/00216208:11320/26:ZALZ4Y9G - isvavai.cz</a>
Result on the web
<a href="https://www.jstage.jst.go.jp/article/jnlp/32/1/32_219/_article/-char/ja/" target="_blank" >https://www.jstage.jst.go.jp/article/jnlp/32/1/32_219/_article/-char/ja/</a>
DOI - Digital Object Identifier
<a href="http://dx.doi.org/10.5715/jnlp.32.219" target="_blank" >10.5715/jnlp.32.219</a>
Alternative languages
Result language
angličtina
Original language name
Data Augmentation for Low-Resource Languages in Multilingual Dependency Parsing
Original language description
UDify (Kondratyuk and Straka 2019) is a multilingual, multi-task parser fine-tuned on mBERT that achieves remarkable performance on high-resource languages. However, on some low-resource languages, its performance saturates early and decreases gradually as training proceeds. To address this issue, this study applies a data augmentation method to improve parsing performance. We conducted experiments on five few-shot and three zero-shot languages to test the effectiveness of this approach. The unlabeled attachment scores were improved on the zero-shot language dependency parsing tasks, with the average score increasing from 55.6% to 59.0%. Meanwhile, dependency parsing tasks in high-resource languages and other Universal Dependencies tasks were almost unaffected. The experimental results demonstrate that the data augmentation method is effective for low-resource languages in multilingual dependency parsing. Furthermore, our experiments confirm that continuously increasing the quantity of synthetic data enhances UDify's performance. This improvement was particularly effective for zero-shot target languages.
Czech name
—
Czech description
—
Classification
Type
J<sub>ost</sub> - Miscellaneous article in a specialist periodical
CEP classification
—
OECD FORD branch
10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)
Result continuities
Project
—
Continuities
—
Others
Publication year
2025
Confidentiality
S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů
Data specific for result type
Name of the periodical
Journal of Natural Language Processing
ISSN
2185-8314
e-ISSN
—
Volume of the periodical
32
Issue of the periodical within the volume
1
Country of publishing house
US - UNITED STATES
Number of pages
33
Pages from-to
219-251
UT code for WoS article
—
EID of the result in the Scopus database
—