Sentence-Ranking-Enhanced Keywords Extraction from Chinese Patents

• Published in: Journal of Information Science and Engineering Vol. 35; no. 3; pp. 651 - 674
• Main Authors: 王志宏(ZHI-HONG WANG), 过弋(YI GUO)
• Format: Journal Article
• Language: English
• Published: Taipei 社團法人中華民國計算語言學學會 01.05.2019; Institute of Information Science, Academia Sinica
• Subjects: Algorithms; Datasets; Information retrieval; Keywords; Performance enhancement; Ranking; key sentences; Chinese patents; sentence-ranking model; human-annotated dataset; keywords extraction
• ISSN: 1016-2364
• DOI: 10.6688/JISE.201905_35(3).0010

Patent keywords, a high-level topic representation of patents, hold an important position in many patent-oriented mining tasks, such as classification, retrieval and translation. However, there are few studies concentrated on keywords extraction for patents in current stage, and neither exist human-annotated gold standard datasets, especially for Chinese patents. This paper introduces a new human-annotated Chinese patent dataset and proposes a sentence-ranking based Term Frequency-Inverse Document Frequency (SR based TF-IDF) algorithm for patent keywords extraction, motivated by the thought of ＂the keywords are in the key sentences＂. In the algorithm, a sentence-ranking model is constructed to filter top-K_S percent sentences from each patent based on a sentence semantic graph and heuristic rules. At last, the proposed algorithm is evaluated with TF-IDF, TextRank, word2vec weighted TextRank and Patent Keyword Extraction Algorithm (PKEA) on the homemade Chinese patent dataset and several standard benchmark datasets. The experimental results testify that our proposed algorithm effectively improves the performance of extracting keywords from Chinese patents.