Research On Aspect-level Sentiment Analysis Method Of Network Reviews Based On Deep Learning

With the rapid development of information technology,posting reviews on the Internet has become an important way for people to express their opinions and convey their experience.Most of these review data contain users' opinions and emotional tendencies towards an entity.By analyzing these review data,users' attitudes can be quickly understood,which can not only provide businesses or customers with richer and more valuable information,but also prevent bad news in time.The development of the incident will then benefit society and the people.However,traditional sentiment analysis can only dig out the user's overall emotional attitude towards a certain entity,and cannot analyze the emotions of different attributes or aspects in the entity.Therefore,in order to conduct a more complete sentiment analysis,it is necessary to discover all aspects of the review data and determine the sentiment and attitude expressed by the review for each aspect.This is the aspect-level sentiment analysis.Aspect-level sentiment analysis is a multi-granularity task in natural language processing.At present,many researchers have carried out research on aspect-level sentiment analysis.In recent years,with the rapid development of deep learning,features that do not require manual construction of feature engineering and have a high degree of automation have been gradually applied in more and more fields,and remarkable effects have also been achieved in aspect level sentiment analysis tasks.However,most of the current research methods focus on the judgment of the affective polarity of the target aspect,while ignoring the importance of aspect item extraction.At the same time,they seldom pay attention to aspect level sentiment analysis for Chinese data.Therefore,in order to better complete the task of aspect-oriented sentiment analysis,this thesis aimed at these problems,carried out research on aspect-oriented sentiment analysis method of network comments based on deep learning,and tried to provide effective solutions.This thesis mainly includes the following aspects:1.Build an aspect-level sentiment analysis framework for network reviews based on deep learning.The framework includes five modules: network review acquisition module,network review preprocessing module,aspect term extraction module,aspect-level sentiment classification module,and testing module.Each module gives the methods of network comment acquisition,network comment preprocessing,aspect term extraction and aspect level sentiment classification.2.Provide a method to obtain network comments using crawler technology.In this method,Selenium and Chromedriver are used to simulate the user directly entering a specified web page with a browser,and the document object model is used to parse the page.At the same time,XPath is used to locate the tag of the web comment body,and the browsing depth is changed in the process of searching and crawling.To get enough comment data.According to the obtained comment data,regular expressions are used to clean up the invalid characters or punctuation in the data,so as to improve the integrity and availability of the data.3.Provide an aspect term extraction method based on the sequence labeling.The sequence labeling uses B,I,O tags to label the original data,and the words or phrases labeled as B or I are regarded as the aspect terms to be extracted.The network first uses the birectional gated recurrent unit to construct the initial contextual semantic representation of the aspect term extraction task.Then,the truncated historical attention is used to obtain the attribute information that has labeled above to guide the attribute labeling below,and position-aware attention is used to calculate the correlation between words while considering the relative positions between words.Finally,these two kinds of feature information are combined to predict the aspect terms of the sequence.At the same time,in view of the extreme imbalance of BIO categories in sequence labeling,this thesis chooses Focal?loss as the loss function,and assigns corresponding weights to the words labeled [B,I,O],so that the network can focus on the content that needs to be extracted.4.An aspect-level sentiment classification method based on deep learning is presented.The method uses both local context information and global sequence information,and uses the local context focus mechanism to notice that different aspects of the text may have different emotional tendencies and reduce the negative effects of the context far away from the target aspect.The multi-head self-attention mechanism is then used to effectively capture the internal structure and contextual dependencies of a sentence,allowing the classifier to interpret different parts of a sentence,while a filter gate is used to delete contextual words that are irrelevant to the current aspect.Finally,the directional self-attention mechanism is used to further model the semantics of aspects and affective words in the sequence,so as to realize the affective polarity analysis of the target aspect.At the same time,the fine-tuning technique is used to train the parameters of the neural network,and the cross entropy loss function is used to avoid the over-fitting problem,so as to improve the training efficiency of the classifier.5.Give comparative experiments and result analysis.First,introduce the public English data set of the international workshop on semantic evaluation and the Chinese data set obtained through crawler technology,development environment and evaluation indicators used in the experiment.Then,through comparative experiments to analyze the effect of the aspect item extraction method and aspect level sentiment classification method given in this article with the current mainstream methods.The experimental results show that,compared with the current mainstream methods,the aspect extraction and classification methods given in this thesis have better performance,and have a certain degree of improvement in each experimental evaluation index,which fully verifies the accuracy and effectiveness of the method in this thesis.Through multiple sets of comparative experiments,it is shown that the aspect term extraction method and classification method given in this article can better improve the efficiency and accuracy of aspect-level sentiment analysis tasks,and also allow users or businesses to have a more comprehensive understanding of online reviews.At the same time,it also allows users or businesses to have a more comprehensive understanding of the fine-grained emotional expressions in network reviews,thereby quickly grasping the sentiments of the people and providing a basis for decision-making by the government or enterprises.