<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0701">
  <Title>part-of-speech tagging of Arabic</Title>
  <Section position="3" start_page="0" end_page="1" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Memory-based learning has been successfully applied to morphological analysis and part-of-speech tagging in Western and Eastern-European languages (van den Bosch and Daelemans, 1999; Daelemans et al., 1996). With the release of the Arabic Treebank by the Linguistic Data Consortium (current version: 3), a large corpus has become available for Arabic that can act as training material for machine-learning algorithms. The data facilitates machine-learned part-of-speech taggers, tokenizers, and shallow parsing units such as chunkers, as exemplified by Diab et al. (2004).</Paragraph>
    <Paragraph position="1"> However, Arabic appears to be a special challenge for data-driven approaches. It is a Semitic language with a non-concatenative morphology. In addition to prefixation and suffixation, inflectional and derivational processes may cause stems to undergo infixational modification in the presence of different syntactic features as well as certain consonants. An Arabic word may be composed of a stem consisting of a consonantal root and a pattern, affixes, and clitics. The affixes include inflectional markers for tense, gender, and number. The clitics may be either attached to the beginning of stems (proclitics) or to the end of stems (enclitics) and include possessive pronouns, pronouns, some prepositions, conjunctions and determiners.</Paragraph>
    <Paragraph position="2"> Arabic verbs, for example, can be conjugated according to one of the traditionally recognized patterns. There are 15 triliteral forms, of which at least 9 are in common. They represent very subtle differences. Within each conjugation pattern, an entire paradigm is found: two tenses (perfect and imperfect), two voices (active and passive) and five moods (indicative, subjunctive, jussive, imperative and energetic). Arabic nouns show a comparably rich and complex morphological structure. The broken plural system, for example, is highly allomorphic: for a given singular pattern, two different plural forms may be equally frequent, and there may be no way to predict which of the two a particular singular will take. For some singulars as many as three further  statistically minor plural patterns are also possible.</Paragraph>
    <Paragraph position="3"> Various ways of accounting for Arabic morphology have been proposed. The type of account of Arabic morphology that is generally accepted by (computational) linguists is that proposed by (Mc-Carthy, 1981). In his proposal, stems are formed by a derivational combination of a root morpheme and a vowel melody. The two are arranged according to canonical patterns. Roots are said to interdigitate with patterns to form stems. For example, the Arabic stem katab (&amp;quot;he wrote&amp;quot;) is composed of the morpheme ktb (&amp;quot;the notion of writing&amp;quot;) and the vowel melody morpheme 'a-a'. The two are integrated according to the pattern CVCVC (C=consonant, V=vowel). This means that word structure in this morphology is not built linearly as is the case in concatenative morphological systems.</Paragraph>
    <Paragraph position="4"> The attempts to model Arabic morphology in a two-level system (Kay's (1987) Finite State Model, Beesley's (1990; 1998) Two-Level Model and Kiraz's (1994) Multi-tape Two-Level Model) reflect McCarthy's separation of levels. It is beyond the scope of this paper to provide a detailed description of these models, but see (Soudi, 2002).</Paragraph>
    <Paragraph position="5"> In this paper, we explore the use of memory-based learning for morphological analysis and part-of-speech (PoS) tagging of written Arabic. The next section summarizes the principles of memory-based learning. The following three sections describe our exploratory work on memory-based morphological analysis and PoS tagging, and integration of the two tasks. The final two sections contain a short discussion of related work and an overall conclusion.</Paragraph>
  </Section>
class="xml-element"></Paper>