Research On Aspect-level Sentiment Classification Based On Dependency Tree And Deep Learning

Aspect-level sentiment classification is a fine-grained sub-task under the domain of text sentiment analysis,which aims to predict the sentiment polarity of each aspect entity in a sentence.With the development of the Internet,various types of social platforms and ecommerce platforms have occupied every corner of people’s daily life.The huge amount of textual information with sentiment tendency that people post on various platforms brings great research value to scientific and commercial fields.By using this textual information,it is not only possible to monitor the direction of public opinion and predict user preferences,but also to further the human understanding of the human brain’s sentiment cognitive system.However,how to efficiently use this sentiment information in conjunction with real-world problems has been a difficult and important focus of research.In recent years,thanks to the rapid development of deep learning techniques,various research areas of natural language processing have made great progress.Similarly,the field of aspect-level sentiment classification has also achieved promising results in recent years.However,the existing schemes do not sufficiently take into account the structural information of the sentences,and the models often get confused and fail to focus on aspect-related sentiment words when dealing with sentences containing multiple aspect words and sentiment words,thus making incorrect sentiment polarity predictions.Therefore,this thesis addresses this drawback by proposing a method that integrates sentence structure information in dependency trees with deep learning techniques to achieve more accurate aspect-level sentiment polarity prediction.The main work and contributions of this thesis are as follows.1.An attention-enhanced graph convolutional network(AEGCN)is proposed.This model mainly solves the noise information problem in traditional graph convolutional network(GCN).AEGCN uses the attention mechanism to calculate the correlation between neighboring nodes in the dependency tree,so that each node in the dependency tree can assign different attention to each node adjacent to it when it is updated,and then distinguish the importance degree of each neighboring node and reduce the noise information propagation.In addition,the model employs a multi-headed attention mechanism to further enhance the robustness of the model by interacting semantic information with structural information.Finally,experimental results on five public English aspect-level sentiment classification datasets show that the AEGCN model effectively improves the classification accuracy by reducing noisy information and further enhances the performance of traditional GCN.2.A triple tag network(TTN)is proposed,which consists of two parts,one is a part-ofspeech-distance network(POSDN)and the other is a relational graph attention network(RGAT).The model mainly explores the impact of tagging information on aspect-level sentiment classification.The dependency tree includes part-of-speech and dependency information in addition to adjacency relationship information between word nodes.In this model,the POSDN uses part-of-speech tag and additional added distance tag to determine the importance level of each word node;the RGAT combines dependency tag with graph attention network to calculate the relevance of neighboring nodes by dependency tags.Finally,guided by the three types of tag information,the feature information captured by the two parts of the network is concatenated and used as the feature representation that the model will eventually use for classification prediction.The experimental results on three public datasets demonstrate the effectiveness of our proposed TTN for the aspect-level sentiment classification task and show that the part-of-speech,dependency distance,and dependency relationship information in the dependency tree can be used as guiding information for the aspect-level sentiment classification task and play a positive role in the prediction of sentiment polarity.3.A dynamic residual graph attention network(DRGAT)is proposed.This model mainly solves the problem of information loss and noisy information in updating node information by existing graph neural network and dependency tree based methods.Firstly,a residual mechanism is introduced into the traditional graph attention network(GAT)to add and normalize the input and output of GAT to overcome the problem of information loss of the original nodes during the network propagation.Secondly,to solve the problem of noisy information,the model introduces four types of tag information into GAT at the same time,and achieves dynamic planning of the attention of each type of tag information based on the traditional GAT.In addition,a dependency tree reshaping method is proposed,which can also alleviate the problem of noisy information in another perspective by pruning off the nodes that are far away from the aspect words in the dependency tree.Finally,the experimental results on three public datasets show that DRGAT can effectively solve the information loss problem and alleviate the problem of noisy information to some extent,and it further improves the performance of GAT.