Aspect-based Sentiment Analysis Of Scenic Spot Comments Based On Deep Learning

The advent of the information age has injected new vitality into various fields.People tend to express views on the Internet,resulting in a large number of comment data.In view of the fact that previous sentiment analysis research focused on the whole sentence or document ignores the different sentiment polarity between various aspect terms and sentence,finegrained sentiment analysis task is proposed.Aspect-based sentiment analysis(ABSA)is a key task in fine-grained sentiment analysis,which aims to predict the sentiment tendency of different aspect terms in a sentence.However,the datasets used in current ABSA are rarely domain-specific,but mostly public.In addition,with the types of online review becoming more and more diverse,aspect-based multi-modal sentiment analysis(ABMSA)task has received much attention.Based on the above background,this thesis applies deep learning methods to mine the tourist’s opinions of the scenic spot.The main work is as follows:1.The technology of web crawler is applied to design a data crawling strategy,and the scrapy crawler framework is employed to obtain tourist’s review data in the field of scenic spot,including scenic spot name,review text,review images,provinces and other related information.According to the aspect terms and data types characteristics,ABSA with text and ABMSA with image and text datasets are constructed,respectively.2.An adaptive dual-channel graph convolution network ABSA model named ADGCN is proposed.Its main idea is to use the syntactic relationship of words to build a global-oriented graph convolutional network to fully learn the features of contextual neighbor nodes.Through pruning,the model reconstructs aspect terms-oriented graph convolutional network and focuses sentiment words around aspect terms to filter the interference of irrelevant words.Then,the attention mechanism is combined to automatically assign the importance scores.The different GCNs features can be adequately fused to extract the most correlated sentiment feature.In this way,the feature is obtained to perform ABSA task.The experimental results prove that the proposed model learns the relationship between aspect terms and sentiment words adequately,and improves the ability of of model classification.3.A dual attention over attention ABMSA(DAOA-ABMSA)is proposed.Firstly,text and images features are extracted using word embedding and Res Net50 network structure in the embedding layer.Then,through the self-attention mechanism,the text and images are encoded,which aims to extract the semantic information inside the complex comment and the multidimensional feature information in the images collection.And for enhancing text and images feature representations,the model combines with attention over attention mechanism to learn hidden information from aspect terms.Finally,text and images features are combined to perform ABMSA tasks.The experimental results verify the effectiveness of the DAOAABMSA and the superiority of multimodality compared with single modality.4.A visual service of the scenic spot comments is designed.The sentiment trend visualization of the scenic spot comments is realized by combining the flask framework and echarts technology.The opinion database is construted based on above models to achieve ABSA and ABMSA online prediction platform,so as to provide decision support for tourist and scenic spot managers.