Classification of Closely Related Sub-dialects of Arabic Using Support-Vector Machines

Classification of Closely Related Sub-dialects of Arabic Using Support-Vector Machines Samantha Wray New York University Abu Dhabi United Arab Emirates samantha.wray@nyu.edu Abstract Colloquial dialects of Arabic can be roughly categorized into five groups based on relatedness and geographic location (Egyptian, North African/Maghrebi, Gulf, Iraqi, and Levantine), but given that all dialects utilize much of the same writing system and share overlapping features and vocabulary, dialect identification and text classification is no trivial task. Furthermore, text classification by dialect is often performed at a coarse-grained level into these five groups or a subset thereof, and there is little work on sub-dialectal classification. The current study utilizes an n-gram based SVM to classify on a fine-grained sub-dialectal level, and compares it to methods used in dialect classification such as vocabulary pruning of shared items across dialects. A test case of the dialect Levantine is presented here, and results of 65% accuracy on a four-way classification experiment to sub-dialects of Levantine (Jordanian, Lebanese, Palestinian and Syrian) are presented and discussed. This paper also examines the possibility of leveraging existing mixed-dialectal resources to determine their sub-dialectal makeup by automatic classification. Keywords: text classification, validation of language resources, language identification 1. Introduction Arabic is a language rich in dialectal variety, with Modern Standard Arabic (MSA) used for official purposes along- side a colloquial dialect local to a speaker’s given region. Although these dialects are typically written throughout in- formal contexts such as social media, they are nonstandard- ized and exhibit spontaneous spelling when compared to the standardized variety. In the text domain, previous work on dialect identification has been performed for purposes ranging from training au- tomatic speech recognition to bootstrapping corpus build- ing efforts. For some of these studies, the focus is on a bi- nary classification between MSA and a colloquial dialect, such as the Egyptian-MSA classification performed by Till- mann et al. (2014), Mansour et al. (2014) and Elfardy and Diab (2013). A more fine-grained classification is performed by (Zaidan and Callison-Burch, 2011), as they crowdsource the human classification of 100,000 sentences of dialectal Arabic, and train language models to classify additional dialectal mate- rial scraped from online newspaper comment sections. The classification categories for this study were MSA, Levan- tine, Egyptian, and Gulf. This 4-way classification achieves an accuracy of 69.4% (Zaidan and Callison-Burch, 2011). However, their accuracy when classifying between just the non-standard dialects Levantine, Egyptian, and Gulf jumps to 83.5%. Typically, dialect classification is performed into these coarse-grained categories. However, work by Malmasi et al. (2015) attempted classification with a finer-grained dis- tinction. They classify text into seven categories: MSA, Tunisian, Egyptian, Jordanian, Syrian, and Palestinian, achieving a 74.3% accuracy for this 6-way classification. However, the training set used to construct the language models was based on materials created by human transla- tion from one dialect to another. This parallel corpus is composed of the above 5 non-standard dialects in addi- tion to MSA from (Bouamor et al., 2014), but the corpus is not spontaneous and was created under …

Content truncated to 500 words