Research On Deep Learning-based Text Sentiment Analysis With Different Granularity

With the rise of online social networks,a large number of users publish posts on the internet to express their sentiment,which include not only the emotional expressions of life and events,but also the user experiences of various aspects of products.Sentiment analysis researches are about the sentiments,opinions,attitudes and emotions in the text.Sentiment analysis tasks have different levels,including sentence-level,target-level and aspect-level.It is necessary to use sentiment analysis with different granularity in different scenarios.Sentiment analysis with different granularity has different problems and challenges due to the characteristics of their own data structures.Firstly,for sentence-level sentiment analysis,the existing methods do not consider how to further mine the data information of the text,or use some output information in the model to further improve the results of existing sentiment analysis models.Secondly,for the target-level sentiment analysis,the sentiment dependencies between different targets are ignored.However,such dependency information between different targets can bring additional valuable information for target-level sentiment classification.Finally,for aspect-level sentiment analysis,existing models still cannot accurately capture the correlation between the words in the sentence and the specific aspects.Confronted with the problems and challenges of sentiment analysis with different granularity,the main research contents of this paper are as follows:?(1)For the task of sentence-level sentiment analysis,by using the information learned from the model to optimize the learning strategy,this paper proposes a dual born again network(DBAN)to further improve the state-of-the-art performance of text sentiment analysis.A dually-born-again network(DBAN)is presented in which the teacher network and the student network are simultaneously trained through an iterative approach.The teacher and the student networks share most of the parameters.All parameters in DBAN are not required to be re-initialized in each training step.Moreover,the DBAN can be further improved by ensemble learning.The proposed DBAN can improve the existing state-of-the-art DNN models in sentiment analysis.Experimental results indicate that DBAN enhances the performances of existing networks.In addition,DBAN outperforms existing born-again network.(2)For the sentence-level sentiment analysis task,by further digging the information in the data,this paper proposes a sample refining mechanism(SR).By studying the relationship between the accuracy of the deep learning model and the probability distribution output of the model,SR estimates the samples that are harmful to model training.Without introducing additional language knowledge and modifying the model structure,SR improves the accuracy of sentiment classification of existing models by reducing the training weight of harmful samples.In addition,this paper introduces ensemble learning to further improve the accuracy of the judgment of harmful samples,thereby further improving the performance of sentence-level sentiment analysis of the original model.Finally,this paper combines SR and DBAN to further optimize the model.It is demonstrated by experiments that combining the SR and DBAN can improve the classification accuracy of sentence-level sentiment analysis.(3)For the target-level sentiment analysis task,in order to effectively model the sentiment dependencies,this paper proposes a novel target-level sentiment classification model based on graph convolutional networks(GCN),called Sentiment Dependencies with Graph Convolutional Networks(SDGCN).SDGCN can effectively capture the sentiment dependencies between multi-targets in one sentence.Our model firstly introduces bidirectional attention mechanism with position encoding to model target-specific representations between each target and its context words,then employs GCN over the attention mechanism to capture the sentiment dependencies between different targets in one sentence.As far as we know,our work is the first to adopt GCN for target-level sentiment classification task.In our method,an aspect is treated as a node,and an edge represents the sentiment dependency relation of two nodes.The proposed approach is evaluated on the Sem Eval 2014 datasets.Experiments show that our model outperforms the stateof-the-art methods.We also conduct experiments to evaluate the effectiveness of GCN module,which indicate that the dependencies between different targets are highly helpful in target-level sentiment classification.(4)For the aspect-level sentiment analysis task,in order to capture the aspect-related context information more accurately,this paper proposes an aspect-level sentiment analysis method based on the Gradual Attention Network(Gra AN).The model is composed of a multi-level gradual attention mechanism.The former-level of the attention mechanism module is relatively smooth,which can notice more words and capture more comprehensive information.The former-level attention mechanism is easier to capture aspect-related information,and generalization ability of the information is also better.The latter-level attention mechanism is connected after the former-level attention mechanism.The latterlevel attention mechanism module is sharper and can capture more accurate local information.Combining multi-level of attention information,the model can grasp both global and local information.In addition,Gra AN adds an auxiliary loss based on the aspect category during the training.By learning the aspect information,the model can learn more aspectrelated information,so that the model can more accurately focus on the words related to the aspect.We conduct extensive experiments on two Sem Eval datasets.The results reveal the essential role of gradual attention mechanism by achieving the state-of-the-art performance.