Detecting Negation and Scope in Chinese Clinical Notes Using Character and Word Embedding
Tian Kang,Shaodian Zhang,Nanfang Xu,Dong Wen,Xingting Zhang,Jianbo Lei
DOI: https://doi.org/10.1016/j.cmpb.2016.11.009
IF: 6.1
2016-01-01
Computer Methods and Programs in Biomedicine
Abstract:BACKGROUND AND OBJECTIVES:Researchers have developed effective methods to index free-text clinical notes into structured database, in which negation detection is a critical but challenging step. In Chinese clinical records, negation detection is particularly challenging because it may depend on upstream Chinese information processing components such as word segmentation [1]. Traditionally, negation detection was carried out mostly using rule-based methods, whose comprehensiveness and portability were usually limited. Our objectives in this paper are to: 1) Construct a large Chinese clinical notes corpus with negation annotated; 2) develop a negation detection tool for Chinese clinical notes; 3) evaluate the performance of character and word embedding features in Chinese clinical natural language processing.METHODS:In this paper, we construct a Chinese clinical corpus consisting of admission and discharge summaries, and propose sequence labeling based systems for negation and scope detection. Our systems rely on features from bag of characters, bag of words, character embedding and word embedding. For scopes, we introduce an additional feature to handle nested scopes with multiple negations.RESULTS:The two annotators reached an agreement of 0.79 measured by Kappa in manual annotation. In cue detection, our systems are able to achieve a performance as high as 99.0% measured by F score, which significantly outperform its rule-based counterpart (79% F). The best system uses word embedding as features, which yields precision of 99.0% and recall of 99.1%. In scope detection, our system is able to achieve a performance of 94.6% measured by F score.CONCLUSIONS:Our study provides a state-of-the-art negation-detecting tool for Chinese clinical free-text notes; Experimental results demonstrate that word embedding is effective in identifying negations, and that nested scopes can be identified effectively by our method.