ABCD Team at SemEval-2023 Task 12: An Ensemble Transformer-based System for African Sentiment Analysis

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 324–330 July 13-14, 2023 ©2023 Association for Computational Linguistics ABCD Team at SemEval-2023 Task 12: An Ensemble Transformer-based System for African Sentiment Analysis Dang Van Thin1,2 and Dai Ba Nguyen1,2 Dang Ba Qui3 and Duong Ngoc Hao1,2 and Ngan Luu-Thuy Nguyen1,2 1University of Information Technology, Ho Chi Minh City, Vietnam 2National University, Ho Chi Minh City, Vietnam 3Nong Lam University, Ho Chi Minh city, Vietnam {thindv,ngannlt,haodn}@uit.edu.vn, 2152914@gm.uit.edu.vn, 21130500@st.hcmuaf.edu.vn Abstract This paper describes the system of the ABCD team for three main tasks in the SemEval-2023 Task 12: AfriSenti-SemEval for Low-resource African Languages using Twitter Dataset. We focus on exploring the performance of ensem- ble architectures based on the soft voting tech- nique and different pre-trained transformer- based language models. The experimental re- sults show that our system has achieved com- petitive performance in some Tracks in Task A: Monolingual Sentiment Analysis, where we rank the Top 3, Top 2, and Top 4 for the Hause, Igbo and Moroccan languages. Besides, our model achieved competitive results and ranked 14th place in Task B (multilingual) setting and 14th and 8th place in Track 17 and Track 18 of Task C (zero-shot) setting. 1 Introduction The AfriSenti-SemEval Shared Task 12 (Muham- mad et al., 2023b) aims at building Sentiment Analysis (SA) systems for 17 African languages, including Hausa, Yoruba, Igbo, Nigerian Pidgin from Nigeria, Amharic, Tigrinya, and Oromo from Ethiopia, Swahili from Kenya and Tanzania, Al- gerian Arabic dialect from Algeria, Kinyarwanda from Rwanda, Twi from Ghana, Mozambique Por- tuguese from Mozambique and Moroccan Ara- bic/Darija from Morocco. This shared task has three main tasks, including two zero-shot tracks, one multilingual track and 12 monolingual tracks. The zero-shot tracks require training two zero-shot models where each model works for only one lan- guage, and each model has training from the 12 languages in the monolingual tracks. The multilin- gual track requires training multilingual SA models that are able to handle all languages. Twelve mono- lingual tracks need training individual monolingual models where each model works for only one lan- guage. The dataset involves tweets labelled with three sentiment classes (positive, negative, neutral) in 14 African languages. Each tweet is annotated by three annotators following the annotation guide- lines in (Mohammad, 2016). They used a form of a majority vote to determine the sentiment of the tweet (Mohammad, 2022; Yimam et al., 2020). All the tweet of the dataset is code-mixed, which can increase the performance of the model. In this paper, we propose an ensemble architec- ture for the AfriSenti-Semeval shared task. Our ensemble architecture is based on the pre-trained transformer-based language models and soft voting technique. In our case, the AfroXLMR, AfriB- ERTa, and LaBSE are employed as the base clas- sifier because these models support African lan- guages in the shared task. The final prediction is combined from the base classifier using the soft voting technique. The rest of the paper is organized as follows. Section 2 provides the …

Content truncated to 500 words