Bhattacharya_Lab at SemEval-2023 Task 12: A Transformer-based Language Model for Sentiment Classification for Low Resource African Languages: Nigerian Pidgin and Yoruba

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1502–1507 July 13-14, 2023 ©2023 Association for Computational Linguistics Bhattacharya_Lab at SemEval-2023 Task 12: A Transformer-based Language Model for Sentiment Classification for Low Resource African Languages: Nigerian Pidgin and Yoruba Nathaniel Hughes1,+, Kevan Baker2,+, Aditya Singh1, Aryavardhan Singh1, Tharalillah Dauda1, and Sutanu Bhattacharya1,* 1Department of Computer Science and Computer Information Systems, Auburn University at Montgomery, Montgomery, AL, USA 2Department of Computer Science, Florida Polytechnic University, Lakeland, FL, USA 1{nhughes4,asingh7,asingh8,tdauda,sbhatta4}@aum.edu 2{kbaker5651}@floridapoly.edu Abstract Sentiment Analysis is an aspect of natural lan- guage processing (NLP) that has been a topic of research. While most studies focus on high- resource languages with an extensive amount of available data, the study on low-resource lan- guages with insufficient data needs attention. To address this issue, we propose a transformer- based method for sentiment analysis for low- resources African languages, Nigerian Pidgin and Yoruba. To evaluate the effectiveness of our multilingual language models for monolin- gual sentiment classification, we participated in the AfriSenti SemEval shared task 2023 com- petition. On the official evaluation set, our group (named as Bhattacharya_Lab) ranked 1 out of 33 participating groups in the Mono- lingual Sentiment Classification task (i.e., Task A) for Nigerian Pidgin (i.e., Track 4), and in the Top 5 among 33 participating groups in the Monolingual Sentiment Classification task for Yoruba (i.e., Track 2) respectively, demon- strating the potential for our transformer-based language models to improve sentiment anal- ysis in low-resource languages. Overall, our study highlights the importance of exploring the potential of NLP in low-resource languages and the impact of transformer-based multilin- gual language models in sentiment analysis for the low-resource African languages, Nigerian Pidgin and Yoruba. 1 Introduction Detecting sentiments or emotions from language is a prominent area of research in natural language processing (NLP) (Medhat et al., 2014; Wankhade et al., 2022). Sentiment analysis (SA) deals with +These authors contributed equally to this work *corresponding author identifying and classifying emotions from textual data (Pang et al., 2002). However, most of the studies focus on high resource languages like English (Ruder, 2020). As such, low resource languages such as African languages with a limited amount of data remain highly underrepresented, making the sentiment classification task for African languages is challenging (Mabokela and Schlippe, 2022). In recent years, there has been increasing interest in developing sentiment classification models for African languages (Muhammad et al., 2022; Ogueji et al., 2021). However, the lack of high-quality labeled datasets, as well as the limited research in this area, limit the progress in this field (Muhammad et al., 2023a). To address this challenge, several research initiatives, such as the SemEval Shared Task (Muhammad et al., 2023b), have been launched to encourage the development of effective sentiment analysis models for African languages. We participated in the SemEval Shared Task 2023, specifically the Monolingual Sentiment Classification task (i.e., Task A) for two low-resource African languages: Yoruba and Nigerian Pidgin. The aim of this task is to develop NLP models capable of accurately …

Content truncated to 500 words