Scholarlink

<p>Sentiment analysis is an essential task in Natural Language Processing (NLP), yet research on low-resource languages remains limited due to the scarcity of annotated corpora. This study introduces the Vaovao Malagasy Sentiment Corpus (VMSC), the first manually annotated dataset for sentiment analysis in Malagasy. The dataset consists of 5,041 sentences collected from Malagasy news articles (2022–2023) and annotated by three native speakers using a strict labeling protocol, achieving a Cohen’s Kappa of 0.96. We benchmark multiple machine learning and deep learning approaches (including Naïve Bayes, Logistic Regression, CNN, BiLSTM, BiGRU) and transformer-based models such as multilingual BERT, DistilBERT, XLM-RoBERTa, and Afro-XLM-R. Experimental results show that Afro-XLM-R achieves the best performance (F1 = 0.8111, accuracy = 0.7980), demonstrating the effectiveness of multilingual pretrained transformers in low-resource settings. The VMSC establishes the first benchmark for Malagasy sentiment analysis and provides a foundation for future research in underrepresented languages. Future work will include expanding to multi-class sentiment, increasing domain coverage, and exploring advanced learning strategies such as domain adaptation and few-shot learning.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Towards robust Malagasy NLP: a novel corpus and evaluation framework for sentiment analysis

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Lorenzo Mamelona,
  • Sitraka Ny Aina Raharivelo,
  • Andriamanarivo Tatiana Miarintsoa,
  • Bright Bediako-Kyeremeh,
  • Benjamin Kwapong Osibo

摘要

Sentiment analysis is an essential task in Natural Language Processing (NLP), yet research on low-resource languages remains limited due to the scarcity of annotated corpora. This study introduces the Vaovao Malagasy Sentiment Corpus (VMSC), the first manually annotated dataset for sentiment analysis in Malagasy. The dataset consists of 5,041 sentences collected from Malagasy news articles (2022–2023) and annotated by three native speakers using a strict labeling protocol, achieving a Cohen’s Kappa of 0.96. We benchmark multiple machine learning and deep learning approaches (including Naïve Bayes, Logistic Regression, CNN, BiLSTM, BiGRU) and transformer-based models such as multilingual BERT, DistilBERT, XLM-RoBERTa, and Afro-XLM-R. Experimental results show that Afro-XLM-R achieves the best performance (F1 = 0.8111, accuracy = 0.7980), demonstrating the effectiveness of multilingual pretrained transformers in low-resource settings. The VMSC establishes the first benchmark for Malagasy sentiment analysis and provides a foundation for future research in underrepresented languages. Future work will include expanding to multi-class sentiment, increasing domain coverage, and exploring advanced learning strategies such as domain adaptation and few-shot learning.