A Research On Deep Learning Based Relation Extration With Improved Hypothesis

Relation extraction is one of the important techniques in information extraction,which aims to extract the relations of labeled entity pairs in sentences.We can not only build knowledge graph on the basic of relation extraction,but also use this technique in text comprehension and machine translation.However,conventional supervised relation extraction methods need massive training dataset to ensure the model can be completely trained,then a lot of experienced experts are required to label the dataset,which is time and labor consuming.Besides,although conventional distant supervision relation extraction can expand the training dataset automatically,these methods will introduce some wrong labelling problems into dataset and decline its quality.The development of deep learning brings new probabilities to natural language processing.Nowadays,techniques like word embedding and neural network in deep learning have great performances on text semantic representation.Therefore,many researchers begin to extract relations at the aspect of deep leaning methods.Howto solve the noise problems in distant supervision data is a burning issue.In this thesis,we improve the basic hypothesis in distant supervision relation extraction methods and design a corresponding deep learning method to fulfill relation extraction task.The whole work can be divided into the following three parts:Firstly,to solve the noise data in distant supervision dataset,we improve the basic hypothesis at the aspect of semantics.In our hypothesis,sentence semantics will be helpful for generating the relation labels of entity pairs.Secondly,in order to increase the quality of whole dataset,we propose a cluster-based distant supervision relation extraction method Clustered DS to relabel the dataset on the basis of improved hypothesis.During the clustering period,the semantics information is conducive to redress the wrong labels in dataset.Finally,we incorporate deep learning into improved hypothesis and propose a Bi-GRU+Clustered DS(Bi-directional Gated Recurrent Unit + Clustered Distant Supervision)model.In this model,the Bi-GRU network is used to extract the hidden semantics of every sentences and the semantics approaching mechanism is designed to achieve better sentence semantic encoding.Experimental results on aligning Freebase knowledge base and the New York Times corpus(NYT)demonstrate that the proposed Clustered DS can settle some wrong labelling and miss labelling problems in distant supervision dataset.Further experiments also show that the proposed Bi-GRU+Clustered DS outperform some of the state-of-art relation extraction methods in precision and recall.