Drug-drug Interaction Extraction Based On Deep Learning Method

With the rapid development of life science and technology,biomedical literature increased exponentially.Massive medical literature is usually written in natural language and stored in an unstructured form.In this literature,there are many potential valuable and rich knowledge.Specially,drug name a wildly studied biomedical entity,concerned by the biomedical research,and as a very important carrier of information extraction from medical literature.If it is possible to extract structured and organized drug information from unstructured text,it will not only be able to expand knowledge corpus like drug dictionaries,but also to provide services to medical researchers and other professional and even promote the development of industrial pharmaceuticals.Thus,the biomedical information extraction is becoming an important research topic.Biomedical information extraction are variety,such as drug name recognition and drug-drug interaction extraction.In other words,to recognize drug name automatically from unstructured biomedical text,then extract the interaction between a pair of drugs.This paper is based on these two issues;the latter two works use deep learning methods.The main contents of the study is as follows.Drug name recognition(DNR)based on enlarging drug name dictionary and conditional random field(CRF)method.Handcraft drug name dictionary are useful for drug name recognition task to improve the performance of the task based machine learning.However,the limited number of hand-building drug name and new drugs are not added in time lead to a limitation when we use the drug dictionary.In this paper,enlarging drug dictionary Drug Bank from the large-scale unstructured biomedical text is based on semi-supervised learning method,which can promote the performance of DNR task based on CRF method.Experiments show that the enlarging drug dictionary effectively improves the performance of drug name recognition.Drug name recognition based on a deep learning method: Long Short-Term Memory and Conditional Random Field model(LSTM-CRF).In DNR task,the stateof-the-art performance is based on machine learning such as CRF method,these methods are heavily relied on handcraft complicated feature and biomedical domain knowledge.In this paper,we constructed bi-directional Long Short-Term Memory(LSTM)and CRF architecture.The inputs of my model is word representations which concatenate character-level word feature to word embedding,the outputs is the label sequence of a sentence.Drug-drug interaction extraction based on a deep learning method: convolutional neural network(CNN),a method comes from deep learning.The best performance of drug-drug interaction extraction based on support vector machine approach;it requires manual construction of complex features and natural language processing tools.In this paper,we use the CNN approach to avoid the complex feature engineering.In our work,input of the proposed approach are word embedding and the position embedding.