The Persian Dependency Treebank Made Universal

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 7078–7087 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 7078 The Persian Dependency Treebank Made Universal Mohammad Sadegh Rasooli1*, Pegah Safari2*, Amirsaeid Moloodi3*, Alireza Nourian4 1Microsoft, Mountain View, CA, USA 2 Faculty of Computer Science and Engineering, Shahid Beheshti University, Tehran, Iran 3 Department of Foreign Languages and Linguistics, Shiraz University, Shiraz, Iran 4 Sobhe, Tehran, Iran 1mrasooli@microsoft.com, 2p_safari@sbu.ac.ir, 3amirsaeid.moloodi@shirazu.ac.ir, 4nourian@sobhe.ir Abstract We describe an automatic method for converting the Persian Dependency Treebank (Rasooli et al., 2013) to Universal Dependencies. This treebank contains 29107 sentences. Our experiments along with manual linguistic analysis show that our data is more compatible with Universal Dependencies than the Uppsala Persian Universal Dependency Tree- bank (Seraji et al., 2016), larger in size and more diverse in vocabulary. Our data brings in labeled attachment F-score of 85.2 in supervised parsing. Also, our delexicalized Persian-to-English parser transfer experiments show that a parsing model trained on our data is ≈2% absolutely more accurate than that of Seraji et al. (2016) in terms of labeled attachment score. Keywords: Universal Dependencies, Persian Dependency Treebank, Automatic Conversion 1. Introduction In recent years, there has been a great deal of inter- est in developing universal dependency treebanks (Mc- Donald et al., 2013; Rosa et al., 2014; Nivre et al., 2020). The main goal of the Universal Dependen- cies project (Nivre et al., 2020) is to develop a con- sistent linguistic annotation scheme in different levels from tokenization to syntactic dependency relations. As a result, the majority of annotation discrepancies disappear, and the resulting dataset facilitates several cross-lingual natural language processing tasks includ- ing part-of-speech transfer (Täckström et al., 2013), syntactic transfer (Naseem et al., 2010; McDonald et al., 2011; Ammar et al., 2016; Zhang et al., 2019), and probing (Tenney et al., 2019; Hewitt and Manning, 2019). Starting with 10 treebanks in 2015, there are 217 treebanks in version 2.9 (November 2021). Persian (aka Farsi) is a pro-drop morphologically rich language with a high degree of free word order and a unique light verb construction (Karimi-Doostan, 2011). Despite its importance, it still suffers from lack of suf- ficient annotated data. The Uppsala Universal tree- bank (Seraji et al., 2016) is currently the only publicly available universal treebank for Persian. It is a valuable resource based on news genre, and has been used as a testbed in previous work (Zeman et al., 2018; Chi et al., 2020). Among other non-universal treebanks, the Persian dependency treebank (PerDT) (Rasooli et al., 2013) is significantly larger than (Seraji et al., 2016) (29K vs. 6K sentences), and its sentences are sampled *Rasooli and Safari equally contributed in the conversion and experimentation process. Rasooli and Moloodi equally contributed in the linguistic design of conversion rules and manual investigation of conversions. .ﺩﺍﺩﺪﻫﺍﻮﺧﻞﯿﻠﻘﺗﻒﺼﻧﻪﺑﺍﺭ�ﻧﺎﻬﺟﺩﺎﺼﺘﻗﺍﺪﺷﺭﺥﺮﻧﻥﺍﺮﯾﺍﻢﯾﺮﺤﺗ . take wiⅼⅼ reⅾuⅽe haⅼf to ra universaⅼ eⅽonoⅿy growth rate Iran sanⅽtion PUNⅭ root SUBJ OBJ ⅯOZ PREⅮEP ⅯOZ ⅯOZ NPOSTⅯOⅮ NPP POSⅮEP NVE root punⅽt nsubj obj …

Content truncated to 500 words