Research On Relation Extraction Method Based On Word Distribution And Deep Residual Network

With the advent of the era of big data,massive amounts of information exist in the network in the form of semi-structures or original texts.How to store the information in a structured form is a problem that needs to be solved.As a structured semantic network,knowledge graph is widely used in natural language processing.Relation extraction can be used to identify the semantic relationship between entity pairs in text,and plays an important role in the semantic understanding of text and the construction and improvement of knowledge graph.The neural network-based relation extraction method is one of the mainstream methods of current relation extraction.It can reduce the complexity of feature engineering in traditional natural language processing,reduce error propagation,and has achieved strong results and generalization in practical applications.Combining the vector representation of word distribution and deep residual network,the research work of relation extraction method is carried out in this thesis.The main work of this thesis is as follows:1.A word distribution model combining word meaning information and word position information is proposed.First,a two-dimensional Gaussian distribution is used to fit a set of word positions for text similarity calculations.Based on this,a word distribution model that combines semantic information and word position information is given.Second,by calculating the similarity between high-frequency words in each cluster and clusters,a sentence-level word distribution vector representation method is proposed.Finally,by calculating the similarity between words and clusters,and combining relational text clustering,a word-level word distribution vector representation method is proposed.The experimental results show that the word-level vector method combined with the word distribution model proposed in this thesis has better effect than the methods in the literature on the distant supervised data Wiki and NYT.2.A relation extraction method based on deep residual network is proposed.First of all,combined with the multi-scale backbone architecture(Res2Net),the hierarchical unit feature layer(Scale)is used in the residual unit structure of the deep network to replace the general convolution layer and expand the receptive field range of each network layer and achieve multi-scale feature representation at different levels of granularity;Secondly,the Squeeze-and-Excitation Networks is integrated to automatically obtain the importance of each feature channel by machine learning in the residual network;Finally,combined with sentence-level attention mechanism,the influence of noise is reduced on the distant supervised dataset,and the relation extraction effect is improved.The experimental results show that the network model proposed in this thesis has achieved better results in distant supervised relation extraction.On the NYT dataset,the method in this thesis improves the Top-K precision of distant supervised relation extraction compared with the methods PCNN + ATT and CNN + RL in the literature.