Biomedical Event Extraction Research Based On Deep Learning

The available biomedical literatures increase at exponential rate in the big data time. The system biologists are urgent to build complex biological relation network. Mining the useful knowledge for them from massive biomedical texts become more and more important. However, traditional relation extraction only extract binary interactions, which cannot satisfy the development of system biology. Consequently, biomedical event extraction which aims to extract more fine-grained and complex biomedical relations between entities has become a hot research topic and has been applied to many domains, such as pathway curation, ontology and semantic network building.Traditional machine learning methods have achieved considerable results in biomedical event extraction. The paper tries to explore further for biomedical event extraction based on the previous research. We mainly study the application of deep learning. We still employ classical procedure of event extraction:trigger identification, argument detection and post process based on rules. First, we employ distributive representation as the feature of words. The dependency-based word embedding is trained based on the contexts of dependency with massive Pubmed abstracts. The embedding can capture functional semantic information for each words. Instead of complicate feature engineering, we build distributive semantic vector based on dependency-based word embedding in the step of trigger identification. Meanwhile, we introduce other useful semantic features including topic, pos and distance. Then, we employ deep learning model to automatically learn senior features which are fed to softmax classifier. In the procedure of argument detection, we employ convolutional neural network to model the sentence based on the dependency path between trigger-entity or trigger- trigger. We also introduce pos, distance and type feature as the supplement of the word embedding and enrich the feature representation. Then, convolution and max-pooling are employed to learn representation for sentences based on the distributed semantic vector. Finally, we process the results of argument detection based on predefined rules and final events are generated through decomposition and composition.We conduct the experiments on MLEE corpora. The model of biomedical event extraction based on deep learning is built based on distributive semantic vector and achieves considerable experiment results. From the experiments, we can also conclude that the other semantic features introduced in trigger identification and argument detection can work well as the supplement of word embedding. At last, we conduct our experiment with similar configuration in other corpus of event extraction for verifying the generalization ability.