大规模生物医学文献医学主题词的高精度自动标注研究

生物医学文献医学主题词的精确标注对于广大生物医学研究人员意义重大，不仅能够帮助研究人员快速找到所需文献，而且为他们进一步挖掘文献，发现新的知识奠定坚实基础。本项目的主要目标是开发大规模生物医学文献医学主题词的高精度自动标注算法，帮助标注人员高效精确标注每年70多万篇生物医学文献，从而更好的服务于广大生物医学研究人员。本课题以生物医学文献医学主题词的预测为研究对象，在机器学习的理论框架指导下，设计基于表示学习的算法获得文献的合适特征表示弥补词袋表示的不足；然后设计基于多标记学习的算法有效处理高达27000多个医学主题词的预测；最后设计基于排序学习的算法有机融合不同表示、预测方法和信息，以提高预测精度。拟解决的科学问题包括：（1）如何结合生物医学文献特点设计合适的文献表示；（2）如何设计考虑医学主题词之间相关性的高效预测算法；（3）排序学习中如何为候选医学主题词设计高区分度特征。

Accurate MeSH (Medical Subject Headings) indexing of biomedical documents is very important for the biomedical researchers. It can not only help the biomedical researchers to find relevant documents, but also facilitate the process of literature mining and knowledge discovery. The main objective is to develop an accurate large scale MeSH indexing algorithm, which can help NLM indexers to efficiently index more than 700,000 new MEDLINE documents every year, and thus provide a better service for the biomedical researchers. The main idea is to develop an accurate indexing algorithm based on machine learning. Firstly, we try to obtain a better representation of biomedical documents based on representation learning; Secondly, we will design an efficient algorithm to handle more than 27,000 MeSH terms by multi-label learning; Finally, based on learning to rank, we will try to improve the prediction performance by integrating different types of representation, predicting methods and MeSH information. Here are three key scientific problems to solve: (1) how to design a better representation by considering the characteristics of biomedical documents; (2) how to design an efficient predication algorithm by considering the dependency of MeSH terms; 3) How to encode the candidate MeSH in the learning to rank framework.