Research On Multi-Channel Neural Network For Relation Extraction

With the advent of the big data era,the amount of data that people need to process is increasing rapidly.How to quickly and effectively extract key information from the vast network has become an urgent problem in the industry.Relation extraction is concerned by researchers because it can extract structured information from large-scale unstructured text.In recent years,scholars have introduced the method based on neural network feature extraction to the task of relation extraction,and have achieved remarkable results.This makes the research of neural network oriented to relation extraction a hot topic.According to different data set labeling methods,relation extraction methods are mainly divided into two categories:full supervision and distant supervision.Among them,the relation extraction method for full supervision mainly has the problem that most neural networks only consider a single branch of information flow,and it is difficult to obtain sufficient semantic features for relation extraction.Relation extraction methods for distant supervision mostly utilize improved relation labels and selective attention-based methods to suppress noisy sentences in bags.Among them,the noise methods are suppressed only from the level of noisy sentences in bags,which has the problem of a single field of view.On the other hand,most distant supervision relation extraction methods only consider a single aspect of feature extraction,lacks effective perception of the global information of the sentence,and cannot filter out noise features from the feature level.To solve above issues,this paper mainly focuses on improving the ability of neural networks to extract semantic features and suppressing noise in distant supervision datasets,and has achieved a series of achievements.The main research content and innovative work of this paper are as follows:(1)In the task of full supervision relation extraction,to solve the issue that most neural networks are difficult to obtain sufficient semantic features,this paper proposes a relation extraction method based on global and local feature-aware network.This method first takes advantage of the self-attention mechanism and recurrent neural network to obtain the correlated sequence features of word.Secondly,a multi-branch feature-aware convolutional neural network is constructed obtain global and local features of correlated sequence,which can avoid the influence between the global and local perceptions.Moreover,the obtained two features are spliced and screened to fully represent the important semantic features of the sentence.Finally,Softmax classifier is utilized to realize relation extraction.Experimental results show that this method performs better than the mainstream relation extraction methods based on convolutional neural networks and recurrent neural networks,and the F₁ of our method reaches 86.1%and 64.9%on the standard Sem Eval-2010 Task 8 and KBP37 datasets,respectively.(2)In the task of distant supervision relation extraction,to solve the issue that the insufficient perception field of most noise reduction methods,this paper proposes a relation extraction method based on multi-level feature refinement.Specifically,for the noisy sentences in the bag,the method first utilizes the feature extraction module to initially extract the features in the bag.Secondly,the feature ensemble module is constructed,and ensemble learning is performed on the initially extracted features in the bag to obtain the ensemble bag features that preliminarily filter out noise.Furthermore,a feature compression module is constructed to compress and condense each ensemble bag features to obtain refined features with essential discrimination capabilities.Finally,the Softmax classifier is combined to realize the relation extraction.Experimental results show that this method performs better than the relation extraction methods based on selective attention and label optimization,and the P@N of this method reaches 85.7%on the standard NYT dataset.(3)In the task of distant supervision relation extraction,to solve the issue that most neural networks lack the effective perception of sentence global information,this paper proposes a distant supervision relation extraction method based on IR-Net.This method improves the quality of representation and suppresses noise features by mining all aspects of related words,sentences and their interactions to clearly obtain the interdependence of all aspects.Specifically,this method first utilizes the response module to modeling the information of words and sentences and obtain the information of many aspects.Secondly,through the interaction module to dynamically summarize various aspects of information,to capture the diversity of global information.Furthermore,an interactive response module is constructed to respond to the dependency relationship between diversity global information and obtain salient global information.Finally,the salient global information is filtered and refined through piecewise max pooling and multi-level feature refinement operations to obtain refined global features for relation prediction.Experimental results show that this method is significantly better than the methods based on selective attention and label optimization in recent years,and the P@N of this method reaches 88.6%on the standard NYT dataset.