Scholarlink

<p>Text preprocessing decisions play a critical role in natural language processing pipelines for morphologically rich, low-resource languages. Stop-word removal is commonly assumed to improve model performance; however, its effect on extractive summarization remains insufficiently examined for Hausa. This study investigates the impact of Hausa-specific stop-word removal on five unsupervised extractive summarization algorithms: TextRank, LexRank, PacSum, RankSum, and HipoRank. Experiments were conducted on a Hausa news corpus using two preprocessing configurations, namely with and without stop-word removal. System outputs were evaluated using ROUGE-1, ROUGE-2, ROUGE-L, and BERTScore-F metrics, and statistical significance was assessed using the Wilcoxon signed-rank test on per-document ROUGE-1 F1 scores. The results show that stop-word removal leads to small, model-dependent variations in automatic evaluation scores; however, none of these differences are statistically significant. Hierarchical and position-aware models exhibit stable performance across preprocessing configurations, while similarity-driven models show minor fluctuations. Human evaluation further indicates that retaining stop words yields higher fluency, coherence, and informativeness ratings for simpler graph-based models, despite negligible differences in automatic metrics. These findings demonstrate that stop-word removal does not provide consistent benefits for Hausa extractive summarization and highlight the importance of combining automatic and human evaluation when assessing preprocessing strategies in low-resource language settings.</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

Investigating the impact of stop words in Hausa extractive summarization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月6日

查看全文
  • Abubakar Salisu Bashir,
  • Abdulkadir Abubakar Bichi

摘要

Text preprocessing decisions play a critical role in natural language processing pipelines for morphologically rich, low-resource languages. Stop-word removal is commonly assumed to improve model performance; however, its effect on extractive summarization remains insufficiently examined for Hausa. This study investigates the impact of Hausa-specific stop-word removal on five unsupervised extractive summarization algorithms: TextRank, LexRank, PacSum, RankSum, and HipoRank. Experiments were conducted on a Hausa news corpus using two preprocessing configurations, namely with and without stop-word removal. System outputs were evaluated using ROUGE-1, ROUGE-2, ROUGE-L, and BERTScore-F metrics, and statistical significance was assessed using the Wilcoxon signed-rank test on per-document ROUGE-1 F1 scores. The results show that stop-word removal leads to small, model-dependent variations in automatic evaluation scores; however, none of these differences are statistically significant. Hierarchical and position-aware models exhibit stable performance across preprocessing configurations, while similarity-driven models show minor fluctuations. Human evaluation further indicates that retaining stop words yields higher fluency, coherence, and informativeness ratings for simpler graph-based models, despite negligible differences in automatic metrics. These findings demonstrate that stop-word removal does not provide consistent benefits for Hausa extractive summarization and highlight the importance of combining automatic and human evaluation when assessing preprocessing strategies in low-resource language settings.