Exploring Machine Translation for code-switching between English and Setswana in South African classrooms

30 Jul 2023 (modified: 07 Dec 2023)DeepLearningIndaba 2023 Conference SubmissionEveryoneRevisionsBibTeX
Keywords: code switching, NLP
TL;DR: MT for CS between English and Setswana in SAn classrooms
Abstract: One of the major challenges of the Department of Education in South Africa is the low numeracy skills amongst South African learners. This study seeks to spotlight the low numeracy challenge encountered in the learning and teaching of mathematics in a classroom where the educator and learners are Setswana native speakers, but use English as the language of learning and teaching. Using English as a language of learning and teaching mathematics to non-native English speaking learners has been stated as one of the reasons why learners perform poorly in mathematics, leading to low numeracy skills. It has been shown that when educators code-switch between English and the native language of the learners to explain mathematical concepts, learners tend to participate more in the classroom and perform better in mathematics. Codeswitching is a topic of interest in Natural Language Processing. Pretrained language models (PLM), such as the mT5, have previously been used in machine translation of code-switched data. In this research, a small corpus of parallel text consisting of monolingual English mathematical text translated into English-Setswana (code-mixed English and Setswana) will be used to fine-tune the mT5 PLM for the purpose of translating mathematical text from English to English/Setswana. In addition, the M2M-100 PLM has been leveraged by African researchers for the machine translation of low-resourced languages. This research aims to fine-tune the M2M-100 PLM using the same corpus and to evaluate its performance on the task of machine translation to potentially aid in the learning and teaching of mathematics in South Africa.
Submission Category: Machine learning algorithms
Submission Number: 48
Loading