Arabic Dialect Identification with a Few Labeled Examples Using Generative Adversarial Networks

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 196–204 November 20–23, 2022. ©2022 Association for Computational Linguistics 196 Arabic Dialect Identification with a Few Labeled Examples Using Generative Adversarial Networks Mahmoud Yusuf Marwan Torki Nagwa El-Makky Computer and Systems Engineering Department Alexandria University Alexandria, Egypt {es-mahmoud.yusuf1217, mtorki, nagwamakky}@alexu.edu.eg Abstract Given the challenges and complexities intro- duced while dealing with Dialect Arabic (DA) variations, Transformer based models, e.g., BERT, outperformed other models in dealing with the DA identification task. However, to fine-tune these models, a large corpus is re- quired. Getting a large number high quality la- beled examples for some Dialect Arabic classes is challenging and time-consuming. In this pa- per, we address the Dialect Arabic Identifica- tion task. We extend the transformer-based models, ARBERT and MARBERT, with unla- beled data in a generative adversarial setting using Semi-Supervised Generative Adversarial Networks (SS-GAN). Our model enabled pro- ducing high-quality embeddings for the Dialect Arabic examples and aided the model to bet- ter generalize for the downstream classification task given few labeled examples. Experimental results showed that our model reached better performance and faster convergence when only a few labeled examples are available. 1 Introduction While Arabic is the first language of most of the Middle East and North Africa (MENA) region, dif- ferent countries have different dialects of Arabic. These Dialect Arabic (DA) forms are all different from the Modern Standard Arabic (MSA). MSA is used in formal writing and speaking situations, like academia and media. In contrast, DA is the language of the street. DA is spoken by people in- formally in their daily conversations and on social media platforms. The task of automatically identifying the dialect of Arabic is beneficial since it contributes to many downstream tasks and applications, such as Speech Recognition and Machine translation. Some Arabic Dialects are very close to each other (e.g. Levantine region dialects such as Lebanese and Syrian). On the other hand, other dialects are significantly different (e.g. Egyptian Class Example English Excuse me, can you take a picture of me? MSA �������� ���� ������� �� � ���� ���� ���� ������ Egyptian ��� ������� ����� � �� �� ������ � Lebanese ������ ���� ����� ����� ������� ��� Moroccan ������� �� ������� � ���� ����� ���� Qatarian ��� ������� ����� � ������ �� Table 1: Comparison between MSA and DA variations for the same sentence and Moroccan dialects) like in Table 1. This simi- larity is affected by the geographic locations of the countries and their respective dialects. Similar dialects are one of the main challenges in the Dialect Identification task. In addition, fur- ther challenges are introduced due to the lack of balanced datasets for DA. Some datasets are imbalanced with few classes dominating the whole dataset. Figure 1 illus- trates the classes distribution in the NADI (Abdul- Mageed et al., 2021b) 2021 dialect dataset. Some other datasets suffer from a …

Content truncated to 500 words