Scholarlink

<p>This paper presents a robust Arabic keyword extraction model designed to address the linguistic complexities of Arabic text. The model follows a multi-phase pipeline, beginning with keyword embeddings generated from Wikidata, followed by candidate identification using Part-of-Speech (POS) rules. Candidate terms are further refined based on contextual relevance, including their occurrence in titles, first paragraphs, and as named entities. A custom ranking algorithm then assigns scores according to factors such as named entity recognition, frequency, and positional salience. To capture the morphological richness of Arabic, evaluation was conducted using a combined metric of "Exact Match + Semantic Similarity + Subwords," where the model achieved a precision of 0.49, recall of 0.33, and F1-score of 0.37 on a test set of 940 Egyptian newspaper articles. These results were obtained through a multi-stage processing framework that integrates linguistic rules, contextual filtering, and ranking strategies, while leveraging parallel computing to efficiently handle large-scale corpora. Parallel processing experiments demonstrated substantial speedup, with optimal efficiency achieved using up to 470 cores, beyond which communication overhead reduced performance gains. Overall, the findings highlight the effectiveness and scalability of the proposed approach in extracting precise and contextually relevant keywords for Arabic NLP applications.</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

Scalable Arabic keyword extraction model with multi-phase analysis using high- performance computing

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Wael S. Afifi,
  • Ibrahim Eldesouky Fattoh,
  • Ahmed Samir Oraby,
  • Hesham ElDeeb,
  • Mohamed Belal,
  • Hoda Baraka,
  • Wael Hassan Gomaa

摘要

This paper presents a robust Arabic keyword extraction model designed to address the linguistic complexities of Arabic text. The model follows a multi-phase pipeline, beginning with keyword embeddings generated from Wikidata, followed by candidate identification using Part-of-Speech (POS) rules. Candidate terms are further refined based on contextual relevance, including their occurrence in titles, first paragraphs, and as named entities. A custom ranking algorithm then assigns scores according to factors such as named entity recognition, frequency, and positional salience. To capture the morphological richness of Arabic, evaluation was conducted using a combined metric of "Exact Match + Semantic Similarity + Subwords," where the model achieved a precision of 0.49, recall of 0.33, and F1-score of 0.37 on a test set of 940 Egyptian newspaper articles. These results were obtained through a multi-stage processing framework that integrates linguistic rules, contextual filtering, and ranking strategies, while leveraging parallel computing to efficiently handle large-scale corpora. Parallel processing experiments demonstrated substantial speedup, with optimal efficiency achieved using up to 470 cores, beyond which communication overhead reduced performance gains. Overall, the findings highlight the effectiveness and scalability of the proposed approach in extracting precise and contextually relevant keywords for Arabic NLP applications.