基于多特征条件随机场的《金匮要略》症状药物信息抽取研究

摘要

目的：结合自然语言处理方法，研究可以有效抽取中医古籍中所含症状和药物文本实体信息的方法。方法以《金匮要略》为例，采用条件随机场（CRF）算法，先将文本进行分词处理，然后以词性、基于键值对的中医诊断标记集作为辅助特征，通过症状-药物 BIO 标签为训练特征来训练出模型，然后利用该模型对测试集文本进行自动标签标注。结果基于多特征 CRF 自动标注的结果准确率达到84.5%，召回率达到70.9%，F测度值达到77.1%。结论运用CRF方法加入词性、中医诊断标记集特征集进行训练得出的多特征模型，能有效提高 CRF 算法对中医古籍的实体抽取能力，生成的模型可用来自动化抽取中医古籍文本的症状药物实体信息。%Objective To find an efficient way to abstract symptoms and medicine information from TCM book Jin Gui Yao Lue through combination of natural language processing method. Methods Taking Jin Gui Yao Lue as an example and by using conditional random fields (CRF), texts were processed according to words, and then part of speech and key assignments based on TCM diagnosis marker group were set as auxiliary features. Symptom-medicine BIO labels were set as the training features to train the model. Then this model was used to conduct automatic labeling to tested texts. Results The accuracy rate of automatic labeling based on multi-feature CRF was 84.5%, recall rate 70.9%, F measure value 77.1%. Conclusion The multi-feature model trained through CRF combined with part of speech and TCM diagnosis marker group can successfully improve abstraction entity information ability from ancient TCM books. The model can be used to automatically abstract symptom and medicine entity information from ancient TCM books.