DN at SemEval-2023 Task 12: Low-Resource Language Text Classification via Multilingual Pretrained Language Model Fine-tuning

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1537–1541 July 13-14, 2023 ©2023 Association for Computational Linguistics DN at SemEval-2023 Task 12: Low-Resource Language Text Classification via Multilingual Pretrained Language Model Fine-tuning Homskiy Daniil homdanil123@gmail.com Maloyan Narek maloyan.narek@gmail.com Abstract In recent years, sentiment analysis has gained significant importance in natural language pro- cessing. However, most existing models and datasets for sentiment analysis are developed for high-resource languages, such as English and Chinese, leaving low-resource languages, particularly African languages, largely unex- plored. The AfriSenti-SemEval 2023 Shared Task 12 aims to fill this gap by evaluating senti- ment analysis models on low-resource African languages. In this paper, we present our solu- tion to the shared task, where we employed dif- ferent multilingual XLM-R models with clas- sification head trained on various data, includ- ing those retrained in African dialects and fine- tuned on target languages. Our team achieved the third-best results in Subtask B, Track 16: Multilingual, demonstrating the effectiveness of our approach. While our model showed rela- tively good results on multilingual data, it per- formed poorly in some languages. Our findings highlight the importance of developing more comprehensive datasets and models for low- resource African languages to advance senti- ment analysis research. We also provided the solution on the github repository. 1 1 Introduction Sentiment analysis, sometimes referred to as opin- ion mining, is a prominent research domain within natural language processing (NLP). Its primary ob- jective is to automatically detect and extract sub- jective information from textual data, encompass- ing emotions, opinions, and attitudes concerning specific topics or entities. Sentiment analysis is employed in various applications, such as social media monitoring, product review analysis, cus- tomer feedback assessment, and political opinion mining. 1https://github.com/Daniil153/SemEval2023_ Task12 Most of the existing sentiment analysis research has concentrated on high-resource languages, in- cluding English and Chinese, while low-resource languages, especially African languages, remain largely unexplored. Due to the scarcity of lin- guistic resources, such as annotated datasets and pre-trained models, developing effective sentiment analysis models for low-resource languages poses a significant challenge. Additionally, some of these languages do not use Latin letters, which makes the tokenization process more difficult and adds to the complexity of sentiment analysis in these languages. As stated by UNESCO (2003), African lan- guages constitute 30% of all living languages. Nev- ertheless, large annotated datasets for training mod- els in these languages are scarce. The AfriSenti- SemEval 2023 competition (Muhammad et al., 2023b) (Muhammad et al., 2023a) aims to inves- tigate models that perform well in low-resource languages. The contest encompasses 14 languages: Hausa, Yoruba, Igbo, Nigerian Pidgin from Nige- ria, Amharic, Xitsonga, Tigrinya, and Oromo from Ethiopia, Swahili from Kenya and Tanzania, Al- gerian Arabic dialect from Algeria, Kinyarwanda from Rwanda, Twi from Ghana, Mozambican Por- tuguese from Mozambique, and Moroccan Ara- bic/Darija from Morocco. Our proposed system utilizes a pre-trained afro- xlmr-large model, which is based on the XLM-R model and trained on 17 African languages and 3 high-resource languages (Alabi et al., 2022). …

Content truncated to 500 words