Improved Keyword Extraction for Persian Academic Texts Using RAKE Algorithm; Case Study: Persian Theses and Dissertations

• Published in: Pizhūhishnāmah-i pardāzish va mudiriyyat-i iṭṭilāʻāt (Online) Vol. 37; no. 1; pp. 197 - 228
• Main Authors: Mehrabi, Elaheh, Mohebi, Azadeh, Ahmadi, Abbas
• Format: Journal Article
• Language: Persian
• Published: Iranian Research Institute for Information and Technology 01.09.2021
• Subjects: keyword extraction; natural language processing; part of speech tagging; persian scientific document; rake algorithm
• DOI: 10.52547/jipm.37.1.197
• ISSN: 2251-8223

Keywords and key phrases are subsets of most relevant words or phrases that summarize contents of a document while they play a critical role in information and document retrieval. Keyword extraction from scientific text is challenging and time-consuming due to the technical and multi-subject nature of the text, while the number of documents requiring keywords is increasing. There are various algorithms and methods developed for automatic keyword extraction. Rapid Automatic Keyword Extraction (RAKE) is a popular algorithm in this domain. RAKE’s decisions are based on the observation that keywords generally contain multiple words and they rarely include stopwords and words with minimum lexical meanings. Candidate keywords are a set of single-word or multi-word sequences selected based on the scores assigned to them by some scoring criteria in RAKE. In this research, a new modified version of RAKE algorithm is proposed in which candidate keyword scoring scheme is improved to increase precision and recall in the keyword extraction process. The proposed algorithm is to cover some of the main weaknesses of RAKE algorithm, especially in Persian scientific documents. To study the weaknesses of RAKE algorithm and evaluating the proposed modified version of RAKE, a set of metadata of Persian theses and dissertations are used. The result of test and evaluation of the proposed algorithm confirm improvement in precision, recall and F-measure. We study effectiveness of RAKE in extracting keywords from Persian texts. We find that RAKE algorithm often extracts long phrases with redundant words on Persian texts, leading to low accuracy. In this paper, we study sources of scoring inefficiency of RAKE algorithm and propose an improved version of RAKE algorithm with a novel scoring mechanism. Our scoring mechanism overcomes some of the weaknesses in RAKE’s original scoring for Persian texts and yields better results. Our evaluations on Persian corpus demonstrate that our improved RAKE algorithm outperforms original RAKE algorithm by extracting more accurate keyword. Our results show that improved RAKE achieves more than 20% higher precision and recall on average compared to original RAKE.