Research And Application Of Aspect-Based Sentiment Analysis Based On Deep Learning

With the digital transformation of the whole society,people are increasingly relying on Internet platforms such as social networks,e-commerce websites,and short video and audio software in their daily lives.There is a wealth of user commentary information stored on these Internet platforms,and mining this information is of great value for marketing,opinion analysis,and decision support.As a sub-task of NLP,text sentiment analysis has been studied by researchers for a long time,but the traditional coarse-grained analysis can no longer meet the growing demand for analysis,and fine-grained aspect-level sentiment analysis has been the focus of research in recent years.In this paper,we propose solutions to the two major problems of how to establish the connection between aspect words and context and how to mine deep syntactic information in aspect-level sentiment analysis,and the main work is as follows.(1)In order to address the problem that existing text sentiment analysis methods cannot obtain various associations between aspect words and context due to the use of a single attention mechanism,this paper proposes an aspect-level sentiment analysis model based on multi-interaction attention graph convolution.First,the model uses a Bi-GRU in the feature extraction layer to obtain the hidden feature vectors of aspectual words and contexts;then,it uses syntactic dependency trees in the graph convolution layer to obtain syntactic structure information in contexts,and uses a multi-interaction attention mechanism to extract semantic features and syntactic features of aspectual words and contexts respectively;finally,it uses a gating mechanism to fuse the semantic features and syntactic features to Finally,a gating mechanism is used to fuse the semantic and syntactic features to determine the aspectual word sentiment polarity.To verify the effectiveness of the model,experiments are conducted on a series of public datasets such as SemEval and Twitter.The experimental results show that the model has significantly improved its effectiveness in aspect-level sentiment analysis.(2)An aspect-level sentiment analysis model based on the pre-training model BERT and reinforced dependency graph convolution is proposed to address the problem that the syntactic dependency information cannot be fully utilized in the model and lacks the ability to uncover deep syntactic information.The model uses the pre-training model BERT as a word embedding layer model to solve the problem of multiple meanings at a time.Also based on the dependency syntactic tree,the dependency relations,dependency types and dependency distances among text words are taken into consideration,so that the model can make full use of all the dependency information and accurately grasp various syntactic structures in text utterances.Moreover,an attention layer focusing on specific aspects is added to the model to focus on contextual information related to aspects.Based on the results of comparative experiments on several publicly available datasets,the model outperforms most recent validated models.(3)Based on the model studied in this paper,an aspect-based sentiment analysis system for hotel reviews is designed and implemented to judge the sentiment polarity of aspect words in hotel review texts,which can help hoteliers improve their service quality based on consumer demand feedback.