Low-resource YouTube comment encoding for Luganda sentiment classification performance

• Published in: Journal of Digital Contents Society Vol. 21; no. 5; pp. 951 - 958
• Main Authors: Ssentumbwe, Abdul Male, Jung, YuChul, Lee, Hyunah, Kim, Byeong Man
• Format: Journal Article
• Language: English
• Published: 한국디지털콘텐츠학회 31.05.2020
• Subjects: 컴퓨터학; Luganda; Low-resource language; 의견 마이닝; Opinion Mining; 유튜브 댓글; YouTube Comments; 저자원 언어; Sentiment Analysis; 감성분석
• ISSN: 1598-2009; 2287-738X
• DOI: 10.9728/dcs.2020.21.5.951

The recent boom in social networks usage has generated some multilingual opinion data for low-resource languages. Luganda is one of the major languages in Uganda, thus it is a low-resource language and Luganda corpora for sentiment analysis especially for YouTube is not easily available. In this paper, we propose assumptions to guide collection of Luganda comments using Luganda YouTube video opinions for sentiment analysis. We evaluate the suitability of our clean YouTube comments (158) dataset for sentiment analysis using selected machine learning and deep learning classification algorithms. Given the low-resource setting, the dataset performs best with Gaussian Naive Bayes for machine learning (55%) and deep learning Multilayer Perceptron sequential model scoring (68.8%) when dataset splitting is at 10% for test set with Luganda comment segmentation. KCI Citation Count: 0