Multi-Dialect Arabic POS Tagging: A CRF Approach

Multi-Dialect Arabic POS Tagging: A CRF Approach Kareem Darwish∗, Hamdy Mubarak∗, Mohamed Eldesouki∗, Ahmed Abdelali∗, Younes Samih†, Randah Alharbi‡, Mohammed Attia⋆, Walid Magdy‡, Laura Kallmeyer† ∗QCRI, †University of Dusseldorf, ‡University of Edinburgh, ⋆Google Inc. {kdarwish,hmubarak,mohamohamed,aabdelali}@hbku.edu.qa, samih@phil.hhu.de, s1581951@sms.ed.ac.uk, attia@google.com, wmagdy@inf.ed.ac.uk, kallmeyer@phil.hhu.de Abstract This paper introduces a new dataset of POS-tagged Arabic tweets in four major dialects along with tagging guidelines. The data, which we are releasing publicly, includes tweets in Egyptian, Levantine, Gulf, and Maghrebi, with 350 tweets for each dialect with appropriate train/test/development splits for 5-fold cross validation. We use a Conditional Random Fields (CRF) sequence labeler to train POS taggers for each dialect and examine the effect of cross and joint dialect training, and give benchmark results for the datasets. Using clitic n-grams, clitic metatypes, and stem templates as features, we were able to train a joint model that can correctly tag four different dialects with an average accuracy of 89.3%. Keywords: Arabic dialects, POS tagging, CRF 1. Introduction Part-of-speech (POS) tagging is important for a variety of applications such as parsing, information extraction, and machine translation. Though much work has focused on POS tagging of Modern Standard Arabic (MSA), work on Dialectal Arabic (DA) POS tagging is rather scant with POS tagged corpora for most dialects being nonexistent or of limited availability. Dialectal POS tagging is becoming increasingly important due to the ubiquity of social media, where users typically write in their dialects to match how they speak in their daily interactions. Dialectal text poses interesting challenges such as lack of spelling standards, pervasiveness of transformative morphological operations, such as word merging and letter substitution or deletion, in addition to lexical borrowing from foreign languages. Existing work on dialectal POS tagging focuses on build- ing resources and tools for each dialect separately (Duh and Kirchhoff, 2005; Habash et al., 2013). The rationale for the separation is that different dialects have different affixes, make different lexical and word ordering choices, and are influenced by different foreign languages. How- ever, performing reliable dialect identification to properly route text to the appropriate POS tagger may be problem- atic, because conventional dialectal identification may lead to results lower than 90% (Darwish et al., 2014). Thus, building a POS tagger that performs reliably across mul- tiple dialects without the need for dialect identification is desirable. In this paper, we present new POS-tagging annotations on a dialectal dataset that is composed of social media text from Twitter for four major Arabic dialects, namely Egyp- tian (EGY), Levantine (LEV), Gulf (GLF), and Maghrebi (MGR) (Eldesouki et al., 2017; Samih et al., 2017a). For each dialect, we tagged 350 tweets using an extended ver- sion of the Farasa tagset (Darwish et al., 2017). We extended the tagset to account for tweet-specific tokens, namely hashtags, user mentions, emoticons and emojis, and URL’s. We created 5-fold partitions for cross-validation with 70/10/20 train/dev/test splits for each dialect. We used the new dataset to train dialectal POS taggers for each di- alect separately to test the effectiveness of the taggers …

Content truncated to 500 words