Aspect-Level Sentiment Analysis Based On Syntax Information And Gating Mechanism

Aspect-level sentiment analysis is a fine-grained sentiment analysis task whose goal is to determine the polarity of the sentiment to which the aspect of the evaluation object(also known as the attribute,both referred to in this paper as the aspect)belongs.For managers,asp ect-level sentiment analysis can provide a more sophisticated view of consumers' attitudes towards goods or services,which can be used as the basis for their improvement of products or services.For consumers,aspect-level sentiment analysis can also provide the attitudes and views expressed by other consumers on goods or services,which can also be used as the basis for making purchasing decisions.Although the machine learning algorithms have achieved good experimental results in aspect-level sentiment analysis tasks,these methods largely rely on the effectiveness of artificial feature construction and require a lot of human labor.Using Long-Short Memory Neural Network(short for LSTM)model,rich text information can be learned from sentences without constructing features manually.By introducing the attention mechanism,the importance of context words to aspect s is obtained,which further enhances the predictive power of the model.However,these LSTMbased neural network models adopt the method of mode ling text contents and aspects respectively.For longer text sequences,there may be a risk of information loss in the process of transmission.The attention mechanism-based model lacks the ability to encode aspect and sentiment features effectively.In addition,with the introduction of attention mechanism,the parameters of neural network model are increased,which may lead to high computational complexity and the risk of overfitting.Therefore,aiming at the model based on Long Short-Term Memory neural network and Attention mechanism,the aspect and sentiment features cannot be encoded effectively in the aspect-level sentiment analysis task,which leads to the problem that the text representation is not reasonable enough.The SIGM model based on syntax information and gating mechanism is proposed.The main structure of the SIGM model includes: word embedding layer,Bi-LSTM layer,Syntax Information layer,convolution layer,Tanh-Relu gating unite layer,pooling layer,output layer and Auto-Encoder structure.?.Word embedding layer and Bi-LSTM layer.Gets the word vector representation of each word in the text from the word embedding matrix.Then,the text represented by the word vector is taken as input,and the hidden output representation of each word with context semantic connection is obtained through the bidirectional LSTM neural network.?.Syntactic information layer.Context words at different distances from the aspect have different effects on the sentiment polarity of the aspect in text.This paper proposed the use of the syntactic information of text to focus on the influence of context words with different distance from aspect on aspect sentiment polarity in syntactic path.?.Tanh-Relu gating unit.In order to model the relationshi p between aspect features and sentiment features effectively,two independent convolutional layers are constructed on the hidden output layer of the bidirectional LSTM neural network and connected to the Tanh and Relu nonlinear function respectively.This structure is called the Tanh-Relu gating unit in this paper.This Tanh-Relu gating unit is used to combine the output results of the two convolutional layers,so that the Tanh-Relu gating unit can selectively extract the sentiment features that match the g iven aspect information,and then judge the sentiment polarity of the aspect.?.Auto-Encoder structure.Aspect terms are usually composed of multiple words or phrases in aspect term sentiment analysis tasks.In the traditional method,the vector represe ntation of these words is usually averaged and then taken as the representation of the aspect term,which may encounter the problem of unreasonable representation of the aspect term information.In order to improve the information representation ability of aspect terms,the auto-encoder structure is constructed in this paper to reasonably represent the aspect terms information,so as to achieve the goal of enhancing the aspect terms feature representation.Finally,in order to verify the predictive performa nce of the SIGM model,this paper used Laptop and Restaurant review dataset for experiment.Experimental results show that the accuracy and F1 value of the SIGM model are better than other comparison models,confirming the effectiveness of the model.Meanwhile,further analysis shows that the SIGM model can predict the sentiment polarity of different aspects in text effectively.