JCT at SemEval-2023 Tasks 12 A and 12B: Sentiment Analysis for Tweets Written in Low-resource African Languages using Various Machine Learning and Deep Learning Methods, Resampling, and HyperParameter Tuning

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 365–378 July 13-14, 2023 ©2023 Association for Computational Linguistics Abstract In this paper, we describe our submissions to the SemEval-2023 contest. We tackled subtask 12 - “AfriSenti-SemEval: Sentiment Analysis for Low-resource African Languages using Twitter Dataset". We developed different models for 12 African languages and a 13th model for a multilingual dataset built from these 12 languages. We applied a wide variety of word and char n-grams based on their tf- idf values, 4 classical machine learning methods, 2 deep learning methods, and 3 oversampling methods. We used 12 sentiment lexicons and applied extensive hyperparameter tuning. 1 Introduction Sentiment analysis (SA), which is the computational determination of the sentiment expressed in a text, has become an increasingly important task in the field of Natural Language Processing (NLP). With the rapid growth of social media platforms, SA has been used to monitor and understand public opinion, classify sentiment, and even mitigate the spread of offensive language. Despite the challenges of SA, such as the variability and subjectivity of sentiment expressions, researchers have been developing and improving SA systems to accurately determine the sentiment expressed in a text. SA is a difficult task for computers and humans. To perform it in High-quality, this task requires understanding the context of the situation, the relevant culture, and in some cases the specific issue or people involved in this situation (Maynard and Greenwood, 2014). The noisy nature of social media texts in general, and on tweeter in particular, makes the analysis task even harder. Therefore, detecting sentiment using supervised machine learning (ML), deep learning (DL) methods and natural language processing (NLP) tools is an interesting and challenging task. SA in low-resource African languages is an even more challenging task due to the limited availability of annotated data, the variability of sentiment expressions, and the lack of NLP tools and resources In this paper, we describe our research and participation in (1) subtask 12-A, for SA in tweets written in 12 languages: Hausa, Yoruba, Amharic, Swahili, Darija, Algerian, Nigerian Pidgin, Igbo, Xitsonga, Twi, Kinyarwandan, and Mozambique Portuguese, and (2) in subtask 12-B for SA in a multilingual tweet corpus, with tweets from all the 12 African languages mentioned above. The full description of task 12 in general and the subtasks, in particular, is given in Muhammad et al. (2023A), and the dataset is described in Muhammad et al. (2023B). The structure of the rest of the paper is as follows. Section 2 introduces a background concerning SA in general and particularly in low- resource African languages, text classification (TC) with imbalanced classes, sentiment lexicons, and hyperparameter tuning. Section 3 describes subtasks 12-A, 12-B, and their datasets. In Sections 4 and 5, we present the submitted models and their experimental results. Section 6 summarizes and suggests ideas for future research. The appendix presents (1) the details of the training, dev, and test sets, (2) our competition best-submitted results, and (3) the sentiment lexicons that we …

Content truncated to 500 words