Trinity at SemEval-2023 Task 12: Sentiment Analysis for Low-resource African Languages using Twitter Dataset

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1161–1165 July 13-14, 2023 ©2023 Association for Computational Linguistics Trinity at SemEval-2023 Task 12: Sentiment Analysis for Low-resource African Languages using Twitter Dataset Shashank Rathi PICT, India shashankrathi2@gmail.com Rahul Tangsali PICT, India rahuul2001@gmail.com Siddhesh Pande PICT, India siddheshspande@gmail.com Aditya Vyawahare PICT, India aditya.vyawahare07@gmail.com Harshwardhan Atkare PICT, India atkareharsh@gmail.com Dipali Kadam PICT, India ddkadam@pict.edu Abstract This paper presents a summary of our findings obtained on sentiment analysis of 3 African languages among the 17 languages mentioned in the shared task. We carried out a sentiment analysis on Hausa, Yoruba, and Swahili. The models used here for the mentioned task were logistic regression, SVM, RandomForest, and mBERT along with a few data-preprocessing and oversampling techniques. The performance of the models used was evaluated by consider- ing weighted average and macro average F1 scores as metrics. The best set of scores ob- tained on the languages Hausa, Yoruba and Swahili are (76.53, 76.55), (74.83, 73.15) and (57.79, 48.59) respectively. 1 Introduction Most languages are becoming digitally accessi- ble as a result of the growing usage of the In- ternet and social media platforms. This enables numerous artificial intelligence (AI) applications that make it possible to do operations like senti- ment analysis, machine translation, and the detec- tion of offensive information. 2,058 or 30 per- cent of all surviving languages, according to UN- ESCO (2003), are African languages. The ma- jority of these languages, however, lack curated datasets for creating these AI applications. The Lacuna Fund is one recent example of a financed program that aims to buck this trend and provide such datasets for African languages. To maximise the use of these datasets, new natural language processing (NLP) techniques must be developed, and research is needed to ascertain whether the present approaches are appropriate in both cases. The shared task covers 17 African languages namely Hausa, Yoruba, Igbo, Nigerian Pidgin from Nigeria, Amharic, Tigrinya, and Oromo from Ethiopia, Swahili from Kenya and Tanza- nia, Algerian Arabic dialect from Algeria, Kin- yarwanda from Rwanda, Twi from Ghana, Mozam- bique Portuguese from Mozambique and Moroc- can Arabic/Darija from Morocco(Muhammad et al., 2023b). The major system strategies used here include us- ing some traditional classifier models like logistic regression, SVM, and RandomForest. They were used through various pipelines. These pipelines also included a vectorizer and over-sampling tools like ROS and SMOTE. One more result per lan- guage was obtained using the mBERT model. All the results obtained were compared and some con- clusions can be drawn based on that analysis. The evaluation metrics used in this shared task were average-macro and average-weighted F1 scores. Since in the shared task, Afrisenti SemEval 2023 we had to deal with African languages it helped us gain familiarity with handling multi-lingual data. We explored and researched different models and data-preprocessing techniques available for such kind of sentiment analysis tasks. 2 Background Here, we have used the dataset provided in the shared task itself. Thus, the input contains tweets …

Content truncated to 500 words