Biomedical Event Extraction Based On Semantic Space And Neural Network

Biomedical literature is the most important way to show and spread the relevant knowledge in the field of biomedicine.With the explosive growth of the literature data,how to quickly and accurately locate the information needed has become an urgent problem to be solved by the researchers in various fields.The main purpose of biomedical information extraction technology is to automatically unlock structured semantic information from unstructured biology and medical texts.The biomedical event extraction is intended to detect the multivariate semantic relationship between fine-grained entities and to show the details of the events in a structured form to the people.In this thesis,biomedical event extraction is a key problem of our research,and it contains two sub-problems: trigger classification and argument detection.The traditional method of detecting the triggers is based on feature engineer,constructing feature vector,and the information used is limited.Therefore,in order to avoid the complexity of human design features and improve the generalization ability of the system,this thesis uses the neural network to automatically learn the feature representation of the candidate instance from the semantic space of the words,and proposes two methods to achieve trigger identification.We propose a method of trigger classification based on convolution neural network,and uses the ability to learn local features of convolution neural network to identify the triggers.Then we propose another method which is based on the bidirectional long short term memory neural network and the conditional random field to aim at trigger classification.We use neural network to learn and integrate context information effectively,and exploit conditional random field to consider global information.We use the method of bidirectional long short term memory neural network and Attention mechanism to carry out biomedical event argument detection.We extract the context feature,distance feature and trigger-argument pair features.Word embedding obtained by the semantic space is used to construct the semantic features,and the bidirectional long short term memory neural network is used to model and study the timing information between triggers and arguments.At the same time,the Attention mechanism is introduced to pay attention to the important features of the candidate instances and give them a greater attention probability.Then we sum the outputs of our model according to the learnt weight values as the final feature vector representation,achieving the argument detection.In this thesis,we experiment on the MLEE corpus.We extract the semantic features of the words from the semantic space which is constructed on the unlabeled data and exploit the neural network to achieve the biomedical event extraction model.The experimental results show that our proposed biomedical event extraction model based on semantic space and neural network is effective.