Research On Fine-grained Sentiment Classification Of Text Based On Deep Learning

Research on machine assisted text analysis follows the rapid development of digital media.Sentiment Analysis(SA)is among the prevalent applications of Natural Language Processing(NLP),a research area that helps machines understand,process and analyze human Language.The traditional methods of sentiment analysis rely on complex feature engineering,which needs a lot of manpower and material resources,and performs poorly across domains.Models based on word embedding have dominated the leaderboard for a long time,but their context-free nature limits their power of expression in complex semantic scenarios.Bidirectional Encoder Representations from Transformers(BERT),among other pre-trained language models,beats existing best results in eleven NLP tasks(including sentence-level sentiment classification)by a large margin,which makes it the new baseline of text representation.Compared with sentence-level sentiment classification,the fine-grained sentiment classification task for a target word(or a specific object)has more technical connotation and higher difficulty.As a more challenging task,fewer applications of BERT have been observed for sentiment classification at the aspect level.How to apply BERT model to the task of sentiment classification for specific target words more concisely and efficiently is the key to improve the accuracy of the model in this field.In order to make the BERT model successfully applied to a particular target word sentiment classification task,we mainly consider four aspects to analyze and improve:(1)The BERT model uses multilayer Transformer network to deeply integrate the target word and its context information.Does the expression of the target word directly generated by the BERT model cover the characteristics of the sentiment association between the target word and the context,and is it enough to distinguish the corresponding polarity of sentiment?(2)BERT model is a representative of dynamic word embedding.Can it further enhance the context perception when the generated word embedding expression is combined with the traditional complex neural network model based on static word embedding expression?(3)Is there any loss in the global information of context for the target words generated by the BERT model?If the global information of sentences is integrated,can the classification accuracy be improved?(4)When multiple objects coexist in a sentence,how do we make use of the inner relation between target words in the sentiment polarity to improve the model classification accuracy?We implement three target-dependent variations of the BERTbase model,with positioned output at the target terms and an optional sentence with the target built in.Experiments on three data collections show that our TD-BERT model achieves new state-of-the-art performance,in comparison to traditional feature engineering methods,embedding-based models and earlier applications of BERT.At the same time,the construction of sentence pair classification task is beneficial to improve the stability and accuracy of the model to a certain extent after the fusion of global features of sentences.However,when BERT model is combined with the complex neural network used for word embedding,the perception of BERT model is not enhanced,sometimes with performance below the vanilla BERT-FC implementation.Existing solutions do not work well when multiple targets coexist in a sentence.The reason is that the existing solution is usually to separate multiple targets and process them separately.If the original sentence has N targets,the original sentence will be repeated for N times,and only one target will be processed each time.To some extent,this approach degenerates the fine-grained sentiment classification task into the sentence-level sentiment classification task.Based on the above considerations,we propose a graph neural network model to model multiple target words appearing simultaneously in sentences based on positional relation.Then,a graph of the sentiment relationship between the target words is constructed by combining the difference of the sentiment polarity between the target words.In addition to the standard object word sentiment classification task,an auxiliary node relation classification task is constructed.Experimental results show that the classification accuracy of the model is improved compared with the previous solutions.Furthermore,the method of dividing multiple objects into isolated individuals has disadvantages,and the multi-task learning model is beneficial to enhance the feature extraction and expression abilities of the model.In the future,we will explore a more detailed and scientific way of composition between target words,as well as the relationship between the influence of the sentiment polarity between target words A directed relationship diagram will be established about the influence of the sentiment tendency between target words.We believe that the classification accuracy of each target word can be improved in the case of multi-objective coexistence,and it may also be a solution with mixed sentiment polarity for a target object.