Supersense Tagging for Arabic: the MT-in-the-Middle Attack

Proceedings of NAACL-HLT 2013, pages 661–667, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Supersense Tagging for Arabic: the MT-in-the-Middle Attack Nathan Schneider∗ Behrang Mohit† Chris Dyer∗ Kemal Oflazer† Noah A. Smith∗ School of Computer Science Carnegie Mellon University ∗Pittsburgh, PA 15213, USA †Doha, Qatar {nschneid@cs.,behrang@,cdyer@cs.,ko@cs.,nasmith@cs.}cmu.edu Abstract We consider the task of tagging Arabic nouns with WordNet supersenses. Three approaches are evaluated. The first uses an expert- crafted but limited-coverage lexicon, Arabic WordNet, and heuristics. The second uses un- supervised sequence modeling. The third and most successful approach uses machine trans- lation to translate the Arabic into English, which is automatically tagged with English supersenses, the results of which are then pro- jected back into Arabic. Analysis shows gains and remaining obstacles in four Wikipedia topical domains. 1 Introduction A taxonomic view of lexical semantics groups word senses/usages into categories of varying granulari- ties. WordNet supersense tags denote coarse seman- tic classes, including personand artifact(for nouns) and motionand weather(for verbs); these categories can be taken as the top level of a taxonomy. Nominal supersense tagging (Ciaramita and Johnson, 2003) is the task of identifying lexical chunks in the sen- tence for common as well as proper nouns, and la- beling each with one of the 25 nominal supersense categories. Figure 1 illustrates two such labelings of an Arabic sentence. Like the narrower problem of named entity recognition, supersense tagging of text holds attraction as a way of inferring representations that move toward language independence. Here we consider the problem of nominal supersense tagging for Arabic, a language with ca. 300 million speak- ers and moderate linguistic resources, including a WordNet (Elkateb et al., 2006), annotated datasets (Maamouri et al., 2004; Hovy et al., 2006), monolin- gual corpora, and large amounts of Arabic-English parallel data. The supervised learning approach that is used in state-of-the-art English supersense taggers (Cia- � �������������� �������� �� ��� � ��� �� �� ���������� ����� ������� Ann-A Gloss Ann-B controls communication manager communication the-windows in attribute configuration relation shape and-layout shape communication windows communication the-applications ‘The window manager controls the configuration and layout of application windows.’ Figure 1: A sentence from the “X Window System” ar- ticle with supersense taggings from two annotators and post hoc English glosses and translation. ramita and Altun, 2006) is problematic for Ara- bic, since there are supersense annotations for only a small amount of Arabic text (65,000 words by Schneider et al., 2012, versus the 360,000 words that are annotated for English). Here, we reserve that corpus for development and evaluation, not training. We explore several approaches in this paper, the most effective of which is to (1) translate the Arabic sentence into English, returning the alignment struc- ture between the source and target, (2) apply En- glish supersense tagging to the target sentence, and (3) heuristically project the tags back to the Arabic sentence across these alignments. This “MT-in-the- middle” approach has also been successfully used for mention detection (Zitouni and Florian, 2008) and coreference resolution (Rahman and …

Content truncated to 500 words