From Constituency to UD-Style Dependency: Building the First Conversion Tool of Turkish

Proceedings of Recent Advances in Natural Language Processing, pages 761–769 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_087 761 From Constituency to UD-Style Dependency: Building the First Conversion Tool of Turkish Aslı Kuzgun Starlang Yazılım Danıs¸manlık asli@starlangyazilim.com O˘guz Kerem Yıldız Ahmet Keles¸o˘glu High School oguz@starlangyazilim.com Neslihan Cesur Starlang Yazılım Danıs¸manlık nesli@starlangyazilim.com B¨us¸ra Mars¸an Starlang Yazılım Danıs¸manlık busra@starlangyazilim.com Arife Bet¨ul Yenice Starlang Yazılım Danıs¸manlık arife@starlangyazilim.com Ezgi Sanıyar Starlang Yazılım Danıs¸manlık ezgi@starlangyazilim.com O˘guzhan Kuyrukc¸u Starlang Yazılım Danıs¸manlık oguzhan@starlangyazilim.com Bilge Nas Arıcan Starlang Yazılım Danıs¸manlık bilge@starlangyazilim.com Olcay Taner Yıldız ¨Ozye˘gin University olcay.yildiz@ozyegin.edu.tr Abstract This paper deliberates on the process of build- ing the first constituency-to-dependency con- version tool of Turkish1. The starting point of this work is a previous study in which 10,000 phrase structure trees were manually trans- formed into Turkish from the original Penn Treebank corpus. Within the scope of this project, these Turkish phrase structure trees were automatically converted into UD-style dependency structures, using both a rule-based algorithm and a machine learning algorithm specific to the requirements of the Turkish lan- guage. The results of both algorithms were compared and the machine learning approach proved to be more accurate than the rule-based algorithm. The output was revised by a team of linguists. The refined versions were taken as gold standard annotations for the evaluation of the algorithms. In addition to its contribu- tion to the UD Project with a large dataset of 10,000 Turkish dependency trees, this project also fulfills the important gap of a Turkish con- version tool, enabling the quick compilation of dependency corpora which can be used for the training of better dependency parsers. 1 Introduction There are two types of annotated treebanks used in natural language processing (NLP) systems: 1https://github.com/StarlangSoftware/StructureConverter constituency treebanks and dependency treebanks. They are used to represent syntactic relations, ar- gument structures, and other hierarchical relations. Constituency treebanks display groups of phrases as trees and dependency treebanks marks head- dependent relations for each item. Today, depen- dency parsers are expected to adapt to new data from a variety of genres. The Universal Depen- dencies (UD) Project (Nivre et al., 2020) provides convenient corpora for these parsers. However, this kind of an adaptation requires a huge amount of data that is not suitable for manual annotation. Thus, the existent constituency treebank corpora are being automatically converted into dependency treebanks as the need for dependency treebanks increased in order to train dependency parsers (Marneffe et al., 2006); (Johansson and Nugues, 2007); (Choi and Palmer, 2012). Constituency to dependency conversion re- quires a big corpus and a coherent annotation framework. UD helps to provide a framework for annotated treebanks to be used in different languages. There have been several attempts made to represent Turkish language in this project and to create a common framework in Turk- ish language treebank studies (Sulubacak et al., 2016); (Oflazer et al., 2003); (C¸ ¨oltekin, 2015) (Kuzgun et al., 2020). However, Turkish language 762 has flexible word order as a consequence of its agglutinative morphology. Such features compli- cates the annotation scheme for Turkish and …

Content truncated to 500 words