Research Of Social Media Relation Extraction Based On Neural Network

Relation extraction is the basic technology of many natural language processing tasks which can provide some structured knowledge semantic understanding and relational reasoning for some downstream tasks such as knowledge Graph and intelligent question answering etc.Some supervised relation extraction tasks depend on large scale datasets which consume lots of manpower and material resources.Although some relation extraction model based on remote supervision can construct lots of training data automatically by aligning knowledge base with unstructured text and reduce the dependency of manual annotation data.But this method introduces lots of noise and false recognition.Therefore,in the field of relation extraction of natural language processing,few shot is involved in researchers' vision.Prototypical network is a method of few shot learning which consists of three key steps: feature extraction,prototype selection and distance measurement.Based on prototypical network.We propose three approaches to improve few shot relation extraction based on two steps: feature extraction and prototype selection.These methods are listed as follow:1)A few shot relation extraction method has been proposed for fused depth convolution neural networks.Compared with ordinary neural networks,deep neural networks contain more hidden layers which can extract text information and location information better from sentence vectors,so the feature vectors of instances are richer.We propose a novel method fused deep convolution neural networks based on few shot relation extraction.This method replaces ordinary single layer convolution neural networks to deep convolution neural networks.Experimental results show that this method can optimize neural networks to some extent.2)Proposes a prototype selection method by fusion attention mechanism.In prototypical network,mean method is used to select prototypes,which is static and fixed.This approach may lose some information on sentence levels and reduce performance of the entire model.We propose two kinds of dynamic prototype selection methods which fused attention mechanism which compensate the shortcomings of prototypical network.This method can select prototypes dynamically and accurately,consider more semantic information,which help to improve accuracy and it is a new strategy for dynamic prototype selection.3)Proposes a Chinese relation extraction method based on fusion depth neural networks and attention mechanisms.Most of existing relation extraction datasets are English based,while Chinese relation extraction datasets are relatively scarce.Firstly we build a Chinese relation extraction dataset.Subsequently we fuse two methods mentioned above,For feature extracting layer,we uses deep convolutional neural network and for prototypes selection,we uses attention mechanism.Finally we use our Chinese dataset to train this model.By using the method,making the relationship extraction work better combine the actual needs in the Chinese domain which reflecting the practical significance of relation extraction in natural language processing field.Based on prototypical network method in few shot relation extraction,our methods make improvements on different aspects of prototypical network and finally makes all innovations integrated.Gradually solve key problems encountered in different development contexts and lay a solid foundation for real production of relation extraction.