Exploring Segmentation Approaches for Neural Machine Translation of Code-Switched Egyptian Arabic-English Text

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3523–3538 May 2-6, 2023 ©2023 Association for Computational Linguistics Exploring Segmentation Approaches for Neural Machine Translation of Code-Switched Egyptian Arabic-English Text Marwa Gaser,1 Manuel Mager,2∗Injy Hamed,3,4 Nizar Habash,4 Slim Abdennadher,1 Ngoc Thang Vu3 1The German University in Cairo, 2AWS AI Labs 3University of Stuttgart, 4New York University Abu Dhabi {marwa.saleh,slim.abdennadher}@guc.edu.eg, magerlm@amazon.com, {injy.hamed,nizar.habash}@nyu.edu, thang.vu@ims.uni-stuttgart.de Abstract Data sparsity is one of the main challenges posed by code-switching (CS), which is fur- ther exacerbated in the case of morphologically rich languages. For the task of machine trans- lation (MT), morphological segmentation has proven successful in alleviating data sparsity in monolingual contexts; however, it has not been investigated for CS settings. In this paper, we study the effectiveness of different segmen- tation approaches on MT performance, cov- ering morphology-based and frequency-based segmentation techniques. We experiment on MT from code-switched Arabic-English to En- glish. We provide detailed analysis, examining a variety of conditions, such as data size and sentences with different degrees of CS. Empir- ical results show that morphology-aware seg- menters perform the best in segmentation tasks but under-perform in MT. Nevertheless, we find that the choice of the segmentation setup to use for MT is highly dependent on the data size. For extreme low-resource scenarios, a combi- nation of frequency and morphology-based seg- mentations is shown to perform the best. For more resourced settings, such a combination does not bring significant improvements over the use of frequency-based segmentation. 1 Introduction Code-switching (CS), i.e. the alternation of lan- guage in text or speech, has been gaining world- wide popularity, due to several reasons, including globalization and immigration. While this has been met with a growing interest in the NLP field to build systems that can handle such mixed input, work on CS machine translation (MT) is still con- sidered in its infancy, where only a few language pairs have been investigated (Sinha and Thakur, 2005; Dhar et al., 2018; Menacer et al., 2019; Xu and Yvon, 2021; Hamed et al., 2022c). ∗Work done while at the University of Stuttgart. CS Sentence ﺔﺣاﺮﺼﺑﺔﺣاﺮﺻ#ب ﻲﻟﺎﺒﺴﻨﻟﺎﺑي#ل#ﺎﺒﺴﻨﻟا#ب عع لا لا it depends لا ع ﻲﻟﺎﺒﺴﻨﻟﺎﺑ ﺔﺣاﺮﺼﺑ situation it depends situation for me it honestly depends on the situation it it it depends depends depend#s honestly for me on the situation situation situation bSrAHp bAlnsbAly E Al bSrAHp b#SrAHp bAlnsbAly b#AlnsbA#ly E E Al Al Translation Word Translation Segmentation Figure 1: An example sentence with code-switching (CS) between English and Egyptian Arabic. The words are contrasted with their segmentations and English translations. Arabic words are paired with their translit- erations in the Buckwalter scheme (Habash et al., 2007). In this work, we focus on the CS Egyptian Ara- bic (EGY)-English (EN) language pair, as we ob- serve its usage is becoming more common. Besides being prevalent amongst Egyptian migrant com- munities, it is also commonly used in Egypt due to the increase in international schooling systems and educational advancements. We identify three main …

Content truncated to 500 words