Is Arabic Part of Speech Tagging Feasible Without Word Segmentation?

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 705–708, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Is Arabic Part of Speech Tagging Feasible Without Word Segmentation? Emad Mohamed, Sandra K¨ubler Indiana University Department of Linguistics Memorial Hall 322 Bloomington, IN 47405 USA {emohamed,skuebler}@indiana.edu Abstract In this paper, we compare two novel methods for part of speech tagging of Arabic without the use of gold standard word segmentation but with the full POS tagset of the Penn Ara- bic Treebank. The first approach uses com- plex tags without any word segmentation, the second approach is segmention-based, using a machine learning segmenter. Surprisingly, word-based POS tagging yields the best re- sults, with a word accuracy of 94.74%. 1 Introduction Arabic is a morphologically rich language, which a word carries not only inflections butin also clitics, such as pronouns, conjunctions, and prepositions. This morphological complexity also has consequences for the part-of-speech (POS) annotation of Arabic: Since words can be com- plex, POS tags refer to segments rather than to whole words. Thus, the word wsyrfEwnhA (in Buckwalter transliteration; engl.: and they will raise it) is assigned the following POS tag: [CONJ+FUTURE PARTICLE+IMPERFECT VERB PREFIX +IMPERFECT VERB+IMPERFECT VERB SUFFIX MAS- CULINE PLURAL 3RD PERSON+OBJECT PRONOUN FEMININE SINGULAR] in the Penn Arabic Treebank (ATB) (Bies and Maamouri, 2003); the boundaries between segments are depicted by + signs. Auto- matic approaches to POS tagging either must assign such complex tags from a large tagset to complete words, or they must segment the word first and then assign POS tags to the segments. Previous approaches (Diab et al., 2004; Habash and Rambow, 2005; van den Bosch et al., 2007; AlGahtani et al., 2009) chose the segmentation approach but concen- trated on POS tagging by using the segmentation provided by the ATB. Additionally, Diab et al. and Habash and Rambow used a reduced tagset. Diab et al. and Habash and Rambow used Support Vector Machines, the former with a standard windowing approach, the latter performing a full morphological analysis before POS tagging. Van den Bosch et al., whose approach is the most similar to ours, used memory-based learning with the full ATB tagset. They report a POS tagging accuracy of 91.5% (93.3% on known words, 66.4% on unknown words). However, they also evaluated on words as defined in the ATB, which differs from written Arabic in the treatment of affixes with syntactic functions (see section 2 for details). AlGahtani et al. used transformation-based learning combined with a morphological analysis for unknown words and words containing clitics. They reached a POS tagging accuracy of 96.9% on ATB1. Surprisingly, their results are lower for the experiment using the whole ATB (96.1%). In this paper, we present two methods for Ara- bic POS tagging that do not require gold stan- dard segmentation but can rather be used for natu- rally occurring Arabic. We investigate two differ- ent approaches: (1) Assigning complete POS tags to whole words, without any segmentation, and …

Content truncated to 500 words