Attention Based Neural Networks For Biological Relation Extraction With Weakly Supervised

Relation extraction is a very important part of natural language processing(NLP).The main content of the relationship extraction task is to obtain the relationship information between the target entities in the corpus,which is actually a multi-classification process.Relation extraction has a wide range of applications in tasks such as "the construction of knowledge map" and "question answering system".In the early stage of relation extraction task,traditional methods such as "conditional random field"have occupied a very "import position.But in the face of increasingly complex data structures and massive amounts of data to be processed,traditional methods are gradually unable to achieve the desired results.In recent years,with the development of deep learning technology,deep learning methods such as convolutional neural networks(CNN)and recurrent neural networks(RNN)are used in relation extraction tasks to obtain more corpus information,and the extraction results are more accurate.At the same time,the application of relationship extraction in biomedical entities is becoming more widespread.The construction of biological knowledge base and the collation of medical data are inseparable from the relationship extraction.In the relationship extraction task,there are problems that the concentration of the target words is insufficient in the process of processing,and the semantic information of the words is not utilized.At the same time,in view of the particularity of biomedical entities,it is necessary to make targeted adjustments to the network model.This paper designs two relation extraction models:Because of the contextual information for long text is easy to be lost,and different sentences or different words have different effects on the relation extraction tasks.This thesis proposes the MAGRU(Multiple Attention GRU)model.The model is based on the GRU(Gated Recurrent Unit),which is one of the variants of the LSTM(Long Short-Term Memory).Moreover,this thesis introduces attention mechanisms on word-layer and sentence-layer.Considering the particularity of biomedical texts,this thesis adjust the attention mechanism to meet the requirements of biomedical entity relationship extraction.The model is compared with the existing better methods on the traditional datasets and biomedical datasets.The experimental results show that the MAGRU model has an advantage of at least l%in F values over existing models.Because of the insufficient use of semantic information in corpus,especially in biomedical texts,semantic information has a greater impact on the relationship extraction effect.Based on the MAGRU model,this thesis also introduces other semantic information such as name entity recognition and part of speech as supplement to the input data.At the same time,the thesis will add semantic information with biomedical background in the biomedical relation extraction experiments,and the attention mechanism will make corresponding adjustments.The model has been compared with the existing methods and the MAGRU model with no semantic information on the traditional dataset and biomedical dataset.The experimental results show that the MAGRU model with semantic information has a 3%improvement in F value compared with other models.