Research On Fine-grained Sentiment Analysis Based On Deep Learning

The Internet generates massive amounts of data all the time,most of which are text data with rich information.Using deep learning technology to mine text data,predict and analyze the sentiment contained in the text has significant commercial and social value.Fine-grained sentiment analysis is an important branch in the field of sentiment analysis.In the research of fine-grained sentiment analysis,a text comment contains multiple evaluation objects(also called aspects).More accurate and multi-dimensional review of text comment data has broad application prospects.It is particularly important for fine-grained sentiment analysis to accurately judge the opinion expressions corresponding to different aspects in a sentence,and to improve the discrimination of the feature vector used to predict the sentiment polarity of the aspect.In current research,recurrent neural networks and attention mechanisms are often used to extract aspect-related contexts,but the extraction accuracy of aspect-related contexts needs to be improved,and there is a lack of contextual information fusion methods.At the same time,the review text often contains some aspects with weak sentiment polarity,which increases the difficulty of the model's sentiment polarity prediction for aspects.In response to the above problems,this paper proposes a fine-grained sentiment classification model Dual BERT that fuses local information and global information,and provides a unified solution based on three aspects: contextual information mining,contextual information fusion,and hard sample mining.The work of this paper is as follows:1.Mining the most relevant local context of an aspect by using a syntactic dependency tree.In order to better establish the connection between the aspect and its corresponding sentiment expression of opinion,the relative semantic distance is obtained by the closest distance between the word nodes in the syntactic dependency tree.The relative semantic distance is used to obtain the weights of different words in the sentence,reducing the influence of irrelevant emotional expressions on the current aspect sentiment polarity prediction,and realizing the mining of aspect local context.2.Use the Bi Affine module to achieve contextual information fusion for local information and global information.In order to accurately obtain the connection between the aspect and the entire sentence and obtain the semantic information of the entire sentence,the sentence and the aspect word are spliced together and input into BERT.At the same time,in order to better integrate the local context information and the global sentiment information of the sentence,the Bi Affine module is used for fusion of local information and global information.3.By artificially constructing positive samples and negative samples to train the feature recognition network,the ability of the backbone network to classify weak sentiment polarity aspect is enhanced.In order to make the feature representation used for predicting sentiment polarity more discriminative,so that some aspects with weak sentiment polarity can be accurately classified,positive samples and negative samples are artificially constructed to train a feature recognition network,which is combined with the backbone network training,so that the feature representation generated by the backbone network is more discriminative.Combining the above three aspects,a sentiment classification model Dual BERT,which integrates local information and global information,is proposed.4.Validated on the Restaurant dataset,Laptop dataset,and Twitter dataset commonly used in fine-grained sentiment analysis,the Dual BERT model proposed in this paper achieved 87.23%,80.72% and 76.3% accuracy,respectively.Meanwhile,the results of ablation experiments show that the three modules proposed in the Dual BERT model are all effective.