Active learning-based information structure analysis of full scientific articles and two applications for biomedical literature review
Active learning-based information structure analysis of full scientific articles and two applications for biomedical literature review
复制标题
DOI：
10.1093/bioinformatics/btt163
复制
发表时间：
2013-06
期刊：
影响因子：
5.8
通讯作者：
Yufan Guo;Ilona Silins;U. Stenius;A. Korhonen
中科院分区：
文献类型：
--
作者：
Yufan Guo;Ilona Silins;U. Stenius;A. Korhonen
动机能够自动分析科学文章的信息结构的技术对于改善生物医学文献的信息访问非常有用。然而，大多数现有的方法依赖于有监督的机器学习（ML）和大量的标记数据，这些数据的开发和应用对于生物医学的不同子领域来说是昂贵的。最近的研究表明，最小的监督是足够的，相当准确的信息结构分析的生物医学摘要。然而，考虑到它们高度的语言和信息复杂性，它对全文是否现实？我们引入并发布了一个新的语料库，其中包含50篇根据论证性分区（AZ）方案注释的生物医学文章，并在此语料库上研究了使用最广泛的ML模型之一-支持向量机（SVM）的主动学习。此外，我们还介绍了两个新的应用程序，使用AZ通过问答和总结来支持生物医学中的真实文献综述。结果我们发现，在500个标记句子（语料库的6%）上训练SVM的主动学习表现惊人，准确率为82%，仅比完全监督学习低2%。在我们的问答任务中，生物医学研究人员从AZ注释的文章中找到相关信息的速度明显快于未注释的文章。在摘要任务中，从特定区域提取的句子比从完整文章的特定部分提取的句子更类似于金标准摘要。这些结果表明，全文信息结构的主动学习确实是现实的，准确率足够高，以支持现实生活中的生物医学文献综述。可用性注释的语料库，我们的AZ分类器和两个新的应用程序可在http://www.cl.cam.ac.uk/yg244/12bioinfo.html
MOTIVATION Techniques that are capable of automatically analyzing the information structure of scientific articles could be highly useful for improving information access to biomedical literature. However, most existing approaches rely on supervised machine learning (ML) and substantial labeled data that are expensive to develop and apply to different sub-fields of biomedicine. Recent research shows that minimal supervision is sufficient for fairly accurate information structure analysis of biomedical abstracts. However, is it realistic for full articles given their high linguistic and informational complexity? We introduce and release a novel corpus of 50 biomedical articles annotated according to the Argumentative Zoning (AZ) scheme, and investigate active learning with one of the most widely used ML models-Support Vector Machines (SVM)-on this corpus. Additionally, we introduce two novel applications that use AZ to support real-life literature review in biomedicine via question answering and summarization. RESULTS We show that active learning with SVM trained on 500 labeled sentences (6% of the corpus) performs surprisingly well with the accuracy of 82%, just 2% lower than fully supervised learning. In our question answering task, biomedical researchers find relevant information significantly faster from AZ-annotated than unannotated articles. In the summarization task, sentences extracted from particular zones are significantly more similar to gold standard summaries than those extracted from particular sections of full articles. These results demonstrate that active learning of full articles' information structure is indeed realistic and the accuracy is high enough to support real-life literature review in biomedicine. AVAILABILITY The annotated corpus, our AZ classifier and the two novel applications are available at http://www.cl.cam.ac.uk/yg244/12bioinfo.html