Using the Semantic-Syntactic Interface for Reliable Arabic Modality Annotation

International Joint Conference on Natural Language Processing, pages 410–418, Nagoya, Japan, 14-18 October 2013. Using the Semantic-Syntactic Interface for Reliable Arabic Modality Annotation Rania Al-Sabbagh Department of Linguistics University of Illinois at Urba- na-Champaign alsabba1@illinois.edu Jana Diesner School of Library and Infor- mation Science University of Illinois at Urba- na-Champaign jdiesner@illinois.edu Roxana Girju Department of Linguistics University of Illinois at Urbana-Champaign girju@illinois.edu Abstract We introduce a novel modality scheme where triggers are words and phrases that convey modality meanings and subcategorize for clauses and verbal phrases. This semantic- syntactic working definition of modality ena- bles us to design practical and replicable anno- tation guidelines and procedures that alleviate some shortcomings of current purely semantic modality annotation schemes and yield high inter-annotator agreement rates. We use this scheme to annotate a tweet-based Arabic cor- pus for modality information. This novel lan- guage resource, being the first, initiates NLP research on Arabic modality. 1 Introduction Modality is the grammaticalized expression of the "speaker's subjective attitudes" (Bybee et al., 1994:176) and "psychological stances" (Mitchell and al-Hassan, 1994:7) towards propositions and events and their factuality status. In NLP applica- tions and domains, modality is considered as one linguistic means to convey and detect attitudes and opinions (Wiebe at al., 2005; Abdul-Mageed and Diab, 2012), commitments and beliefs (Diab et al., 2009), power relations (Prabhakaran et al., 2012), uncertainties and speculations (Szarvas et al., 2008; Matsuyoshi et al., 2010). We herein present the first work on Arabic modality annotation, which is part of a larger research project to (1) automatically identify modality triggers (i.e. words and phrases convey- ing modality meanings), holders (i.e. modality experiencers) and scopes (i.e. the range of lin- guistic constituents modified by the modality triggers), and (2) automatically detect power re- lations among participants in the social network of Twitter by using this modality information. Despite solid work on Arabic modality in the- oretical linguistics (Mitchell and al-Hassan, 1994; Brustad, 2000; Moshref, 2012), there are no Arabic corpora annotated for modality, not even the widely used Penn Arabic Treebank. However, there is a plethora of work and anno- tated corpora for modality in other languages, including English (Saurí et al, 2006; Baker et al., 2010; Prabhakaran et al., 2012; Rubinstein et al., 2013), Portuguese (Hendrickx et al., 2010; Avila and Mello, 2013), Japanese (Matsuyoshi et al. 2010) and Chinese (Cui and Chi, 2013). Arabic modality annotation involves multiple challenges. First, the paradigm of Arabic modali- ty triggers is complex as it includes auxiliaries, lexical verbs, nominals and particles - like many other languages as well. Second, triggers can be lexically and/or semantically ambiguous: a lexi- cally-ambiguous trigger is a word or phrase that may or may not convey a modality meaning based on context. A semantically-ambiguous trigger is a word or phrase that may convey two or more modality meanings. Third, implicit scopes are common in Arabic and annotators have to be made aware of them. Fourth, Arabic word order flexibility allows triggers - especially adverbials - to occur in the …

Content truncated to 500 words