Towards accurate dependency parsing for Galician with limited resources

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11320%2F25%3AL39SNLPG" target="_blank" >RIV/00216208:11320/25:L39SNLPG - isvavai.cz</a>

Výsledek na webu

<a href="https://www.scopus.com/inward/record.uri?eid=2-s2.0-85206510805&doi=10.26342%2f2024-73-18&partnerID=40&md5=0fa04a9f64eb9d360809cbc16f8c0cb2" target="_blank" >https://www.scopus.com/inward/record.uri?eid=2-s2.0-85206510805&doi=10.26342%2f2024-73-18&partnerID=40&md5=0fa04a9f64eb9d360809cbc16f8c0cb2</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.26342/2024-73-18" target="_blank" >10.26342/2024-73-18</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Towards accurate dependency parsing for Galician with limited resources

Popis výsledku v původním jazyce

Automatic syntactic parsing is a fundamental aspect within NLP. However, effective parsing tools necessitate extensive and high-quality annotated treebanks for satisfactory performance. Consequently, the parsing quality for low-resource languages such as Galician remains inadequate. In this context, the present study explores several approaches to improve the automatic syntactic analysis of Galician using the UD framework. Through experimental endeavors, we analyze the quality of the model incrementing the size of the initial training corpus by adding data from Galician PUD treebank. Additionally, we explore the benefits of incorporating contextualized vector representations by comparing the use of various BERT models. Lastly, we assess the impact of integrating cross-lingual training data from similar varieties, analyzing the models’ performance across used treebanks. Our findings underscore (1) the positive correlation between augmented training data and enhanced model performance across used treebanks; (2) superior performance of monolingual BERT models compared to their multilingual analogues; (3) improvement of overall model performance across utilized treebanks by incorporation of cross-lingual data. © 2024 Sociedad Española para el Procesamiento del Lenguaje Natural.

Název v anglickém jazyce

Towards accurate dependency parsing for Galician with limited resources

Popis výsledku anglicky

Automatic syntactic parsing is a fundamental aspect within NLP. However, effective parsing tools necessitate extensive and high-quality annotated treebanks for satisfactory performance. Consequently, the parsing quality for low-resource languages such as Galician remains inadequate. In this context, the present study explores several approaches to improve the automatic syntactic analysis of Galician using the UD framework. Through experimental endeavors, we analyze the quality of the model incrementing the size of the initial training corpus by adding data from Galician PUD treebank. Additionally, we explore the benefits of incorporating contextualized vector representations by comparing the use of various BERT models. Lastly, we assess the impact of integrating cross-lingual training data from similar varieties, analyzing the models’ performance across used treebanks. Our findings underscore (1) the positive correlation between augmented training data and enhanced model performance across used treebanks; (2) superior performance of monolingual BERT models compared to their multilingual analogues; (3) improvement of overall model performance across utilized treebanks by incorporation of cross-lingual data. © 2024 Sociedad Española para el Procesamiento del Lenguaje Natural.

Druh

J_SC - Článek v periodiku v databázi SCOPUS

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

—

Návaznosti

—

Rok uplatnění

2024

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název periodika

Procesamiento del Lenguaje Natural

ISSN

1135-5948

e-ISSN

—

Svazek periodika

2024

Číslo periodika v rámci svazku

73

Stát vydavatele periodika

US - Spojené státy americké

Počet stran výsledku

11

Strana od-do

247-257

Kód UT WoS článku

—

EID výsledku v databázi Scopus

2-s2.0-85206510805