Improved Arabic Dialect Classification with Social Media Data

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2118–2126, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Improved Arabic Dialect Classification with Social Media Data Fei Huang Facebook Inc. Menlo Park, CA feihuang@fb.com Abstract Arabic dialect classification has been an important and challenging problem for Arabic language processing, especially for social media text analysis and machine translation. In this paper we propose an approach to improving Arabic dialect classification with semi-supervised learn- ing: multiple classifiers are trained with weakly supervised, strongly supervised, and unsupervised data. Their combination yields significant and consistent improve- ment on two different test sets. The dialect classification accuracy is improved by 5% over the strongly supervised classifier and 20% over the weakly supervised classifier. Furthermore, when applying the improved dialect classifier to build a Modern Stan- dard Arabic (MSA) language model (LM), the new model size is reduced by 70% while the English-Arabic translation qual- ity is improved by 0.6 BLEU point. 1 Introduction As more and more users share increasing amount of information on various social media platforms (Facebook, Twitter, etc.), text analysis for so- cial media language is getting more important and challenging. When people share their sto- ries, opinions, post comments or tweets on so- cial media platforms, they frequently use collo- quial languages, which are more similar to spo- ken languages. In addition to typical natural lan- guage processing problems, the informal nature of social media languages presents additional chal- lenges, such as frequent spelling errors, improper casing, internet slang, spontaneity, dis-fluency and ungrammatical utterances (Eisenstein, 2014). Di- alect classification and dialect-specific processing are extra challenges for languages such as Arabic and Chinese. Considering Arabic as an example: there are big differences between MSA and various dialec- tal Arabic: MSA is the standardized and literary variety of Arabic used in writing and in most for- mal speech.1 It is widely used in government pro- ceedings, newspapers and product manuals. Many research and linguistic resources for Arabic nat- ural language processing are based on MSA. For example, most existing Arabic-English bilingual data are MSA-English parallel sentences. The di- alect Arabic has more varieties: 5 major dialects are spoken in different regions of the Arab world: Egyptian, Gulf, Iraqi, Levantine and Maghrebi (Zaidan and Callison-Burch, 2011). These dialects differ in morphologies, grammatical cases, vocab- ularies and verb conjugations. These differences call for dialect-specific processing and modeling when building Arabic automatic speech recogni- tion (ASR) systems or machine translation (MT) systems. Therefore, identification and classifica- tion of Arabic text is fundamental for building so- cial media Arabic speech and language processing systems. In order to build better MT systems between Arabic and English, we first analyze the distri- bution of different Arabic dialects appearing on a very large scale social media platform, as well as their effect on Arabic-English machine translation. We propose several methods to improve the dialect classification accuracy by training models with distant supervision: a weakly supervised model is trained with data whose labels are …

Content truncated to 500 words