A Second Wave of UD Hebrew Treebanking and Cross-Domain Parsing

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4331 - 4344 December 7-11, 2022 ©2022 Association for Computational Linguistics A Second Wave of UD Hebrew Treebanking and Cross-Domain Parsing Amir Zeldes Georgetown University amir.zeldes@georgetown.edu Nick Howell IAHLT nlhowell@gmail.com Noam Ordan IAHLT noam.ordan@gmail.com Yifat Ben Moshe IAHLT yifat1811@gmail.com Abstract Foundational Hebrew NLP tasks such as seg- mentation, tagging and parsing, have relied to date on various versions of the Hebrew Tree- bank (HTB, Sima’an et al. 2001). However, the data in HTB, a single-source newswire corpus, is now over 30 years old, and does not cover many aspects of contemporary Hebrew on the web. This paper presents a new, freely available UD treebank of Hebrew stratified from a range of topics selected from Hebrew Wikipedia. In addition to introducing the corpus and eval- uating the quality of its annotations, we de- ploy automatic validation tools based on grew (Guillaume, 2021), and conduct the first cross- domain parsing experiments in Hebrew. We obtain new state-of-the-art (SOTA) results on UD NLP tasks, using a combination of the lat- est language modelling and some incremental improvements to existing transformer based ap- proaches. We also release a new version of the UD HTB matching annotation scheme updates from our new corpus. 1 Introduction Treebanks (TBs) form a fundamental resource for NLP and computational linguistics research: they provide high quality annotated data for tokeniza- tion, sentence splitting, POS tagging and syntac- tic/semantic relation extraction. For Morphologi- cally Rich Languages (MRLs, Seddah et al., 2014), high quality morphosyntactically annotated data is particularly crucial, since basic search on a word- matching level cannot work before morphological segmentation has occurred. In languages such as Hebrew, where vowels are not well represented in the script, the need for reliable morphosyntactic disambiguation is particularly strong, since string- level ambiguity is very frequent and high. This is demonstrated by the following frequently cited ex- ample from Adler and Elhadad (2006), which has a large number of possible analyses for just a four character sequence (note Hebrew is right-to-left). (1) �לצב ⟨b.cl.m⟩ be.cil.am - in.shadow.their ⟨b.clm⟩ (be./b.a.)celem - in.(a/the).image ⟨b.clm⟩ (be./b.a.)calam - in.(a/the).photographer ⟨bcl.m⟩ bcal.am - onion.their ⟨bclm⟩ becelem - Betzelem (organization) When such sequences contain multiple sub- tokens (e.g. ‘in’ and ‘the’ within ‘in.the.image’), we follow Universal Dependencies1 terminology in referring to the larger unit as a Multi-Word Token (MWT), and the sub-parts, each of which carries a separate part-of-speech, as tokens. Although a treebank for Hebrew was already cre- ated by the MILA Institute (Sima’an et al. 2001, hence HTB), subsequently converted to dependen- cies (Goldberg and Elhadad, 2009) and finally to the more recent Universal Dependencies frame- work (see Sade et al. 2018, hence UD-HTB), two major concerns motivate the current work in cre- ating a new UD treebank for Hebrew. The first is the age of the data: HTB texts were taken from 1990–91 issues of a single newspaper, Ha’aretz, as illustrated in (2), describing the introduction of computers to an office. (2) ba-axrona …

Content truncated to 500 words