SemEval-2007 Task 18: Arabic Semantic Labeling

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 93–98, Prague, June 2007. c⃝2007 Association for Computational Linguistics Semeval 2007 Task 18: Arabic Semantic Labeling Mona Diab Columbia University mdiab@cs.columbia.edu Christiane Fellbaum Princeton University fellbaum@clarity.princeton.edu Musa Alkhalifa University of Barcelona musa@thera-clic.com Aous Mansouri University of Colorado, Boulder aous.mansouri@colorado.edu Sabri Elkateb University of Manchester Sabri.Elkateb@manchester.ac.uk Martha Palmer University of Colorado, Boulder martha.palmer@colorado.edu Abstract In this paper, we present the details of the Arabic Semantic Labeling task. We describe some of the features of Arabic that are rele- vant for the task. The task comprises two subtasks: Arabic word sense disambiguation and Arabic semantic role labeling. The task focuses on modern standard Arabic. 1 Introduction Recent years have witnessed a surge in available re- sources for the Arabic language.1 The computa- tional linguistics community is just about starting to exploit these resources toward several interesting scientific and engineering goals. The Arabic lan- guage is interesting from a computational linguistic perspective. It is significantly different from English hence creating a challenge for existing technology to be easily portable to Arabic. The Arabic language is inherently complex due to its rich morphology and relative free word order. Moreover, with the exis- tence of several interesting varieties, the spoken ver- naculars, we are witnessing the emergence of written dialectal Arabic everyday on the web, however there are no set standards for these varieties. We have seen many successful strides towards functional systems for Arabic enabling technolo- gies, but we are yet to read about large Arabic NLP applications such as Machine Translation and Infor- mation Extraction that are on par with performance on the English language. The problem is not the ex- istence of data, but rather the existence of data an- notated with the relevant level of information that 1Author 1 is supported by DARPA contract Contract No. HR0011-06-C-0023. Authors 2, 3 and 4 are supported by the US Central Intelligence Service. is useful for NLP. This task attempts a step towards the goal of creating resources that could be useful for such applications. In this task, we presented practitioners in the field with challenge of labeling Arabic text with seman- tic labels. The labels constitute two levels of gran- ularity: sense labels and semantic role labels. We specifically chose data that overlapped such that we would have the same data annotated for different types of semantics, lexical and structural. The over- all task of Arabic Semantic Labeling was subdivided into 4 sub-tasks: Arabic word sense disambiguation (AWSD), English to Arabic WSD task (EAWSD), argument detection within the context of semantic role labeling, and argument semantic role classifica- tion. Such a set of tasks would not have been feasible without the existence of several crucial resources: the Arabic Treebank (ATB) (Maamouri et al., 2004), the Arabic WordNet (AWN) (Elkateb et al., 2006), and the Pilot Arabic Propbank (APB).2 This paper is laid out as follows: Section 2 will describe some facts about the Arabic language; Sec- tion 3 will present the overall …

Content truncated to 500 words