ASMA: A System for Automatic Segmentation and Morpho-Syntactic Disambiguation of Modern Standard Arabic

Proceedings of Recent Advances in Natural Language Processing, pages 1–8, Hissar, Bulgaria, 7-13 September 2013. ASMA: A System for Automatic Segmentation and Morpho-Syntactic Disambiguation of Modern Standard Arabic Muhammad Abdul-Mageed Indiana University Bloomington, IN, USA mabdulma@indiana.edu Mona Diab George Washington University Washington DC, USA mtdiab@email.gwu.edu Sandra K¨ubler Indiana University Bloomington, IN, USA skuebler@indiana.edu Abstract In this paper, we present ASMA, a fast and efficient system for automatic seg- mentation and fine grained part of speech (POS) tagging of Modern Standard Ara- bic (MSA). ASMA performs segmenta- tion both of agglutinative and of inflec- tional morphological boundaries within a word. In this work, we compare ASMA to two state of the art suites of MSA tools: AMIRA 2.1 (Diab et al., 2007; Diab, 2009) and MADA+TOKAN 3.2. (Habash et al., 2009). ASMA achieves comparable results to these two systems’ state-of-the- art performance. ASMA yields an accu- racy of 98.34% for segmentation, and an accuracy of 96.26% for POS tagging with a rich tagset and 97.59% accuracy with an extremely reduced tagset. 1 Introduction Arabic raises various challenges to natural lan- guage processing (NLP): Arabic is a morpholog- ically rich language (Tsarfaty et al., 2010), where significant information concerning syntactic units is expressed at the word level, which makes part of speech (POS) tagging a challenge since it involves morpho-syntactic disambiguation, including fea- tures like voice, number, gender (Diab, 2007; Diab et al., 2007; Habash et al., 2009). We address the problem of full morpho- syntactic disambiguation of words in context. We devise a system, ASMA, that performs both in- flectional morpheme segmentation and aggluti- native clitic segmentation. For example, given a surface word in context such as ���������� �������� �� (wabiHasanaAtihim, Eng. ‘and by their virtues’)1, 1For Arabic examples, we use both the Arabic script and the Buckwalter Arabic transliteration scheme (Buckwalter, 2004). ASMA provides the following segmentation: �� ��� �� �� �� ��� �� ��� wa bi Hasan aAti him, with the pre- fixal clitics �� ��� (wa bi, Eng. ‘and’ ‘by’), the stem �� �� �� (Hasan), the inflection morpheme ��� �� (aAti), and the suffixal pronominal morpheme ��� (him). ASMA then assigns each one of these resulting morphemes a POS tag. For an explanation of Ara- bic morphology, cf. section 2. The most successful approaches to date that render this level of morphological segmentation (addressing both inflectional as well as aggluti- native boundaries) typically rely on employing a morphological analyzer in the process (Habash et al., 2009). We show that it is possible to efficiently perform full morpho-syntactic disam- biguation employing language-independent meth- ods that are not based on a morphological ana- lyzer. Our motivation is that dependence on a mor- phological analyzer comes at the cost of develop- ment since such an analyzer is generally based on manually written rules and an extensive lexicon. ASMA performs both inflectional morpheme segmentation and agglutinative clitic segmenta- tion, as well as fine grained POS tagging of Mod- ern Standard Arabic (MSA). In ASMA, a seg- ment is a stem, an …

Content truncated to 500 words