Chemical-Disease Relation Recognition Based On Biomedical Literature Mining

Recognition of chemical-disease relations is an important task in biomedical lit-erature mining.Chemical-disease relations can improve search results of biomedical search engine,shorten drug development cycle and reduce drug development costs.There are two subtasks in chemical-disease relation recognition,which are biomedi-cal named entity normalization and chemical-disease relation extraction.Normaliza-tion solves the problem that biomedical names are rich,varied and hard to identify in literature.The main objective of normalization is to map biomedical names in litera-ture to standard biomedical concepts.Current solutions include character matching and relation matrix learning.Character matching method is limited by the size of normal-ization vocabulary,and relation matrix learning method is insensitive to new words.Chemical-disease relation extraction needs to extract relations between chemical and disease based on literature context.Current solutions include co-occurence frequency counting,pattern matching and statistical machine learning.Co-occurence frequency counting method has low precision,and pattern matching or statiscal machine learning methods need human-designed patterns or features which require human efforts and may be not comprehensive or accurate.Following are main works and contributions of this dissertation on the two subtasks:1.Biomedical named entity normalization based on semantic matchingThis dissertation proposes a biomedical named entity normalizaion algorithm based on semantic matching.It excavates words' semantic similarity by their contexts' simi-larity and matches biomedical names by measuring semantic similarity.This disserta-tion first collects lots of unlabelled biomedical contexts and constructs word vectors in semantic space.And then it uses deep networks measuring distance between biomed-ical names in semantic space.A ranking-based loss function and stochastic gradient descent algorithm are used here to train the model.The experiment shows that the pro-posed method gets an accuracy of 83.5%on NCBI test set which is better than traditional methods.2.Chemical-disease relation extraction based on convolutional neural networkThis dissertation proposes a chemical-disease relation extraction algorithm based on convolutional neural network.The algorithm focuses on local key features to extract chemical-disease relation.It uses unsupervisedly trained word vectors and relative lo-cation information as word features,and concats word features togather forming a sen-tence feature map.The convolutional neural network captures local key features from the feature map by convolution and pooling operations.The experiment shows that the proposed method gets an F-score of 50.67%on CDR test set.Compared to traditional methods,the proposed method needs less human efforts while it has stronger covering capacity.