Research On Aspect-level Sentiment Analysis Method Basesed On Graph Convolutional Netwok

With the amount of user-generated information on the Web,identifying the sentiment polarity of the given aspect provides more complete and in-depth results for businesses and customers.Aspect-level sentiment analysis has become one of the important tasks in natural language processing tasks.In recent years,with the wide application of deep learning,great achievements have been made in the aspect-level sentiment analysis task.The aspect-level sentiment analysis task is the understanding task of natural language.Therefore,the judgment of the aspect-level sentiment polarity is not only related to the context content,the local content of the specific aspect,but also the structure of the text.The sentiment class of aspect words may be different in different text structures.However,the existing deep learning methods ignore the influence of text structure features on the aspect-level sentiment analysis.In this thesis,the given text is transformed into the text graph and the given aspect is regarded as a partial region of the graph.The sentiment classification for the aspect is regarded as the local region classification of the graph.The graph convolutional network is applied to the text graph to capture the structural features of the text,improving the performance of the aspect-level sentiment analysis task.However,the natural language understanding task is different from the local area classification task of the general graph.This thesis focuses on the two main task text representations and target representations of aspect-level sentiment analysis,and studies the methods of using graph structures and graph convolutional networks to obtain relevant feature representations for aspect sentiment analysis.The main research contents include:(1)Research on aspect-level sentiment analysis methods based on graph convolutional neural networks and local representations.In order to better capture the structure and syntactic feature information of text,syntactic information and cooccurrence information are used to construct a text graph.The proposed method uses graph convolution network to obtain text representation and structured attention mechanism to obtain aspect representation.Experimental results show that this method can effectively improve the performance of the system.(2)Research on the aspect-level sentiment analysis method based on the attention mechanism of graph structure.A text graph attention model GAM is proposed,which is different from the traditional graph attention mechanism.It divides passing message in the graph convolution network into two parts: adjacency information and target attention information.It improved text representation ability for the aspect based sentiment analysis.Experimental results show that graph convolutional networks using GAM attention mechanism are more suitable for aspect-level sentiment analysis tasks.(3)Research on memory graph convolutional neural network.In the process of long text analysis,more graph convolution layers are needed to obtain the feature information of the text.However,with the increase of the layers of graph convolution network,the representation ability of nodes decreases.In this thesis,a memory graph convolutional network model RGCN is proposed.To improved nodes representation ability,this model introduces a memory network into the graph convolutional network,so that the update information of the nodes is no longer limited to the adjacent nodes.The experimental results show that the aspect-level sentiment analysis framework based on the RGCN can effectively improve the text representation ability of long texts.(4)Research on local region representation method based on local segmentation algorithm.The relationship between words in long texts is complex,and filtering noise information is important for aspect level sentiment analysis.This thesis proposes a local area representation method using the Nibble algorithm.It uses the Nibble algorithm to segment and filter irrelevant information,and uses the local clustering sequence algorithm to obtain the target representation.Experimental results show that the aspect-level sentiment analysis model based on this method can effectively improve the target representation ability of long texts.In summary,the research of this thesis aims to capture the structural information in the text by using graph structure and graph convolutional network.We propose models based on graph convolutional network from different views to make it suitable for aspect-level sentiment analysis task,and verify on the public datasets.The experiments show that the proposed methods can effectively improve the performance of the aspectlevel sentiment analysis task.The research in this thesis has provided a new idea for this task and contributed to exploring new heights in the field of aspect-level sentiment analysis.