Seals_Lab at SemEval-2023 Task 12: Sentiment Analysis for Low-resource African Languages, Hausa and Igbo

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1508–1517 July 13-14, 2023 ©2023 Association for Computational Linguistics Seals_Lab at SemEval-2023 Task 12: Sentiment Analysis for Low-resource African Languages, Hausa and Igbo Nilanjana Raychawdhary1, Amit Das1, Gerry Dozier1, and Cheryl D. Seals1 1Department of Computer Science and Software Engineering, Auburn University, AL, USA 1{nzr0044,azd0123,doziegv,sealscd}@auburn.edu Abstract One of the most extensively researched appli- cations in natural language processing (NLP) is sentiment analysis. While the majority of the study focuses on high-resource languages (e.g., English), this research will focus on low- resource African languages namely Igbo and Hausa. The annotated tweets of both languages have a significant number of code-mixed tweets. The curated datasets necessary to build com- plex AI applications are not available for the majority of African languages. To optimize the use of such datasets, research is needed to deter- mine the viability of present NLP procedures as well as the development of novel techniques. This paper outlines our efforts to develop a sentiment analysis (for positive and negative as well as neutral) system for tweets from the Hausa, and Igbo languages. Sentiment analy- sis can computationally analyze and discover sentiments in a text or document. We worked on the first thorough compilation of AfriSenti- SemEval 2023 Shared Task 12 Twitter datasets that are human-annotated for the most widely spoken languages in Nigeria, such as Hausa and Igbo. Here we trained the modern pre- trained language model AfriBERTa large on the AfriSenti-SemEval Shared Task 12 Twit- ter dataset to create sentiment classification. In particular, the results demonstrate that our model trained on AfriSenti-SemEval Shared Task 12 datasets and produced with an F1 score of 80.85% for Hausa and 80.82% for Igbo languages on the sentiment analysis test. In AfriSenti-SemEval 2023 shared task 12 (Task A), we consistently ranked top 10 by achieving a mean F1 score of more than 80% for both the Hausa and Igbo languages. Keywords: low-resource languages, sentiment analysis, AfriBERTa large model, twitter cor- pus, natural language processing. 1 Introduction Natural language processing (NLP) is a method that allows computers to intelligently and effectively analyze, and comprehend meaning from human language (Lopez and Kalita, 2017). In NLP, trained language models have taken the highest position, displaying outstanding performance on a variety of NLP tasks (Ogueji et al., 2021) because of their capacity to recognize and learn intricate linguistic correlations and patterns in natural language data. Sentiment analysis (SA) (Muhammad et al., 2023a) is a method for computationally identifying and classifying opinions especially to determine if the author has positive, negative, or neutral emo- tions about a specific topic. Furthermore, it has been shown that models like XLM-R (Conneau et al., 2020) and mBERT (Devlin et al., 2018) gen- eralize well to a variety of languages. These models are known to be challenging to utilize since training data is typically lacking for low-resource languages (Muhammad et al., 2022). It is difficult to adapt the majority of high- resource language models successfully and ap- proaches to …

Content truncated to 500 words