基于引文全文本的医学领域突破性文献识别研究*

王 雪 杨雪梅 林紫洛 关陟昊 唐小利

(北京协和医学院/中国医学科学院 医学信息研究所 北京 100005)

Research on Identification of Medical Breakthrough Articles Based on Citing Sentences

Wang Xue Yang Xuemei Lin Ziluo Guan Zhihao Tang Xiaoli

(Peking Union Medical College/Chinese Academy of Medical Sciences, Beijing 100005)

Abstract：[Purpose/Significance]From the citing perspective of the academic community, based on the citing sentences, and word frequency statistics, deep learning, etc., the paper explores the text features that represent breakthrough evaluations in the citing sentences and builds an automatic recognition model to identify potential breakthrough articles. [Method/Process] The authors selected the key publications of the Nobel Prize winners in Physiology or Medicine and the representative articles of Science Breakthrough of the Year as the gold standard breakthrough articles collection and obtained their citation sentences. Word frequency statistics combined with manual screening were carried out to obtain common words that characterize breakthrough evaluation. The authors manually labeled the citing sentences and used BERT and BIOBERT models for training to form automatic recognition models, and finally selected the cancer field for empirical analysis. [Result/Conclusion] The results show that there is obvious textual characteristics when evaluating literatures with great breakthrough value. Compared with the BERT model, the recognition ability of BIOBERT model was improved, with F1 value of 0.84. The automatic recognition model based on citing sentences can accurately identify the literature with important academic value and realize early recognition and early evaluation to a certain extent.

Keywords：citing sentences;deep learning;breakthrough articles;automatic identification;text classification;textual characteristics

科学的重大进展在很大程度上取决于突破性的科学发现[1]。目前，世界各国政府和科研资助机构已对优先发展主题领域或前沿领域投入大量资源和资金支持，特别是突破性研究和变革性研究，以期促进科学发展。早在2007年，美国国家科学委员会就提议加强对变革性研究的支持[2]。近年来，美国国立卫生研究院专门为变革性研究设立资金资助(Transformative Research Award)，旨在支持“具有创造或推翻基本范式潜力的异常创新和/或非常规研究项目”[3]。我国在十九大报告中提出加快建设创新型国家，指出要瞄准世界科技前沿，强化基础研究，实现前瞻性基础研究、引领性原创成果重大突破，其中一个重要方面就是加快推进基础研究领域突破性、颠覆性创新的遴选和培育[4]。尽管各国都已逐步建立资助政策和研究基础设施以激励和促进突破性研究的发展，但最初选择和资助的提案是否最终会成为一项突破性研究仍然是一个重要且不可避免的挑战。

本研究尝试从科学共同体的定性描述出发，结合深度学习等方法，实现从海量文献中自动识别潜在突破性文献。潜在突破性文献的识别，可以为科研人员的研究方向以及资助机构的重点资助方向提供参考，促进潜在突破点、突破性技术的大力发展，同时帮助科研人员准确锁定研究领域中的突破性进展；引用语句体现着学术共同体在引用原始文献时的评价态度，基于引用语句识别潜在突破性文献，丰富了传统学术评价的内容，为基于引用语句实现论文评价提供了新的运用实践。……