Scholarlink

<p>The study presents a transformer-based approach for sentiment classification in the Yoruba language using the multilingual Bidirectional Encoder Representations from Transformers (mBERT) model. Yoruba, a tonal and morphologically rich language, presents unique challenges for computational modeling due to its diacritical orthography and limited digital resources. A manually annotated corpus of 2,000 Yoruba movie reviews was developed, containing balanced positive and negative sentiments. Preprocessing involved Unicode normalization to preserve tonal diacritics and maintain orthographic integrity, avoiding unidecode stripping. The mBERT model, fine-tuned using the Hugging Face Transformers and PyTorch framework, was evaluated against traditional machine learning and deep learning baselines including Naïve Bayes, Support Vector Machine, Long Short-Term Memory (LSTM), and Recurrent Neural Network (RNN). Experimental results revealed that the diacritic-preserving mBERT model achieved 89.4% accuracy, outperforming all baseline models. The inclusion of Unicode normalization improved classification accuracy by 4.5% compared with the diacritic-stripped variant. Macro-averaged metrics confirmed balanced model performance across sentiment classes, while grouped movie-based splitting validated robustness by eliminating data leakage. The findings demonstrate that transformer-based multilingual models can effectively capture contextual and tonal nuances in low-resource African languages. The study highlights the importance of diacritic-sensitive preprocessing and transfer learning in advancing natural language understanding for Yoruba and other underrepresented African languages.</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

Sentiment classification in the Yoruba language using multilingual Bidirectional Encoder Representations from Transformers (mBERT)

  • 期刊论文

  • 发表日期：

    2025年12月12日

查看全文
  • Akinbowale Nathaniel Babatunde,
  • Jumoke Falilat Ajao,
  • Abdulrahman Abdullahi Kayode,
  • Ronke Seyi Babatunde,
  • Salman Olatunji Isiaka,
  • Roseline Oluwaseun Ogundokun,
  • Cheng-Chi Lee

摘要

The study presents a transformer-based approach for sentiment classification in the Yoruba language using the multilingual Bidirectional Encoder Representations from Transformers (mBERT) model. Yoruba, a tonal and morphologically rich language, presents unique challenges for computational modeling due to its diacritical orthography and limited digital resources. A manually annotated corpus of 2,000 Yoruba movie reviews was developed, containing balanced positive and negative sentiments. Preprocessing involved Unicode normalization to preserve tonal diacritics and maintain orthographic integrity, avoiding unidecode stripping. The mBERT model, fine-tuned using the Hugging Face Transformers and PyTorch framework, was evaluated against traditional machine learning and deep learning baselines including Naïve Bayes, Support Vector Machine, Long Short-Term Memory (LSTM), and Recurrent Neural Network (RNN). Experimental results revealed that the diacritic-preserving mBERT model achieved 89.4% accuracy, outperforming all baseline models. The inclusion of Unicode normalization improved classification accuracy by 4.5% compared with the diacritic-stripped variant. Macro-averaged metrics confirmed balanced model performance across sentiment classes, while grouped movie-based splitting validated robustness by eliminating data leakage. The findings demonstrate that transformer-based multilingual models can effectively capture contextual and tonal nuances in low-resource African languages. The study highlights the importance of diacritic-sensitive preprocessing and transfer learning in advancing natural language understanding for Yoruba and other underrepresented African languages.