医学信息领域人工智能技术的主题漂移与未来展望——基于JCR 26本医学信息期刊文本的命名实体识别  

Topic Drift and Future Prospect of Artificial Intelligence Technology in Medical Information Field——Named Entity Recognition of 26 Medical Information Journals Based on JCR

作　　者：徐璐璐[1,2,3] 杨嘉乐 康乐乐 Xu Lulu;Yang Jiale;Kang Lele(Library,Nantong University,Nantong 226019,China;School of Information Management,Nanjing University,Nanjing 210023,China;Jiangsu Key Laboratory of Data Engineering&Knowledge Service,Nanjing 210023,China;School of Information Science and Technology,Nantong University,Nantong 226019,China)

机构地区：[1]南通大学图书馆,江苏南通226019 [2]南京大学信息管理学院,江苏南京210023 [3]江苏省数据工程与知识服务重点实验室,江苏南京210023 [4]南通大学信息科学技术学院,江苏南通226019

出　　处：《现代情报》2022年第10期163-176,共14页Journal of Modern Information

基　　金：江苏省高校哲学社会科学研究一般项目“医学信息智能化发展的主题漂移与趋势分析”(项目编号:2020SJA1591)。

摘　　要：[目的/意义]在过去数十年中,医学信息研究领域被人工智能技术的重构。为厘清人工智能技术发展对医学信息研究领域带来的影响,本研究采用命名实体对医学信息领域人工智能技术进行识别,深入揭示其主题漂移特征与演化趋势,并提出3点未来展望。[方法/过程]研究中首先采集了JCR中26本医学信息期刊题录信息,而后利用Vosviewer可视化分析人工智能技术的总体分布,在此基础上采用3种深度学习模型对人工智能技术进行命名实体识别和对比,最后分5个时间段梳理其主题漂移并提出3点展望。[结果/结论]Vosviewer可视化显示20年来人工智能技术在医学信息领域占据重要地位;3种深度学习模型对比发现,基于Attention的Bi LSTM-CRF模型的命名实体识别结果最优,F1值提高到88.40%;在5个时间段内,医学信息领域人工智能主流技术以高、中频词为代表围绕着传统型技术且相对稳定,分支技术以低频词为代表则出现深度学习等复杂性技术且随时间有所改变,并呈现直觉(经验发掘)→支持(深入理解)→策略(强化分析)→后推理(支撑决策)→前推理(提前预测);即整体进入较为理性和务实状态,尚缺爆发性变革但确有一定程度变化的主题漂移演化脉络。对此,本文从技术、应用和并行层面提出3点未来展望,以期加强对人工智能在处理医学信息上优、缺点的认知,为更精准地挖掘多源数据提供优质医学诊断具有理论和现实意义。[Purpose/Significance]In the past decades, the research field of medical information has been reconstructed by Artificial intelligence technology(AI).In order to clarify the impact of the development of AI in the field of medical information, this study uses named entity recognition technology to identify related technologies, and deeply reveals the characteristics and trends of thematic changes, puts forward three prospects for the future.[Method/Process]In the study, the informations of 26 medical information journals in JCR were collected, and then Vosviewer was used to visually analyze the overall distribution of AI.On this basis, three deep learning models were used to identify and compare the named entities of AI,and finally the development venation was sorted out in five time period.[Result/Conclusion]During the past 20 years, AI has occupied an important position in the field of medical information by Vosviewer;The comparison of three deep learning models shows that the named entity recognition of Bi LSTM-CRF-Attention model is the best, and the F1 value is increased to 88.40%;In the five time periods, The mainstream technologies of AI in the fild of medical information are represented by high and medium frequency words and relatively stable, around the traditional technology, while the branching AI technologies of low frequency words are represented as complex technologies such as deep learning and change over time.In general, the application of AI presents a clear development process: intuition(experience discovery)→support(in-depth understanding)→strategy(enhanced analysis)→post-reasoning(supporting decision)→pre-reasoning(prediction in advance).That is, the whole has entered a more rational and pragmatic state, but there is still a lack of explosive change and also a certain degree of change in the theme drift evolution.This study puts forward three future prospects from the aspects of technology, application and parallel, which strengthens the cognition of the advantages and disadvantages of

关 键 词：医学信息 人工智能技术 命名实体 主题漂移 BERT模型 双向长短期记忆网络 条件随机场 注意力机制 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]