Research On Aspect-level Sentiment Classification Based On Deep Learning

Aspect-level sentiment classification is one of the important research directions of natural language processing,which aims to accurately distinguish the emotional polarity of the target entity in a sentence.Due to the emotional polarity of the target entity depends on some emotional feature words that express the target emotion in the sentence,by excavating the semantic and emotional relationship between the target entity and the sentence,we can find the emotional feature words related to the target entity.Although the aspect-level sentiment classification model has made great progress in the past few decades,there are still a series of problems,such as not fully excavating the relationship between the target entity and emotional feature words,and accurately distinguishing the target emotional polarity in sentences that containing multiple targets.it is still a rather challenging task.In order to accurately judge the emotional polarity of the target entity in the sentence,by studying the current methods in the field of natural language processing,we propose some aspect-level sentiment classification models based on deep learning,the main contents are as follows:(1)A aspect-feature Fusion with Graph Convolutional Network(AFGCN)model is proposed.Firstly,the sentence is input into the word embedding layer,and the context semantic information is ecoded in the Bi-LSTM layer,and obtain the hidden layer representation of the sentence,then,according to the different distances from the target entity of the word in the sentence,the hidden layer representation of the word is multiplied by the corresponding position weight,so as to achieve the effect that the word with a longer distance has less influence.since the target is an entity noun,its emotional tendency is determined by the emotional feature words in the sentence,it does not contain emotional polarity,so first set the target to zero,and then use the graph convolution network,the hidden layer representation of the words that have a dependency syntactic relationship with the target is filled into the target,then use the attention mechanism to filter the key information in the hidden layer representation of the sentence,Finally,it is spliced with the pooled target to get the final result.By combining the key information selected by graph convolution network and attention mechanism,a more accurate emotional feature representation is constructed,thus the classification accuracy is improved.(2)A model which model Multi-aspect Dependencies with Graph Convolutional Network(MDGCN)is proposed.Firstly,model encodes the semantic information to the input sentence,and then encodes the context semantic to the target through the attention mechanism.In order to effectively model the dependency relationship between multiple targets in one sentence,then,constructing multi-aspects dependencies graph from the dependency syntax tree,and employs GCN over the multi-aspects dependencies graph to model the dependencies between multiaspects in one sentence.Finally,sentiment classification is preformed using the target representation generated by the GCN.Through the construction of multi-aspects dependence graph,the dependency relationship between targets is obtained,which makes up for the limitation that a single target is limited by the related words on the dependency syntax tree and can not find more effective emotional feature words,thus improving the classification accuracy.(3)Based on the AFGCN network model,this paper combines the pre-training language model and the improved dependency syntax tree.A model which based on AFGCN over Improved Dependency Tree and BERT(IBGCN)is proposed.Firstly,according to the dependency syntax tree of the sentence,taking the target as the root,the aspect-oriented dependency tree(ADT)is redesigned.A unique dependency syntax tree is generated for each target,thus making full use of the characteristics of the target.Then,in order to enhance the expressive ability of the model and have different word vector representations in specific context,use BERT as the word vector,so that each word in the sentence has a more accurate and contextual hidden layer representation in the high-dimensional space,so as to improve the classification accuracy.