Research And Application Of Aspect Based Sentiment Analysis Method Based On Knowledge Enhancement

Sentiment analysis is one of the most important branches of natural language processing and can be of great use to business decisions by analyzing the sentiment information in user texts.Sentiment analysis tasks can be divided into chapter-level sentiment analysis,sentence-level sentiment analysis and aspect-level sentiment analysis,depending on the granularity.As the times have changed,aspect-level sentiment analysis is gaining more attention in sentiment analysis in order to extract more sentiment information from text data.Aspect-level sentiment analysis aims to extract aspectual items and determine their sentiment polarity based on a given sentence,and is usually divided into two sub-tasks,namely aspectual item extraction and aspectual sentiment classification.For the aspect extraction task,previous studies have typically extracted only aspect items or extracted aspect and opinion items separately without identifying the relationship between them.For aspect sentiment classification,the current widely used pre-trained language models lead to poor classification results due to the difference between their training goals and the goals of aspect-level sentiment analysis.In addition,external knowledge such as knowledge graphs are widely used in natural language processing models to recognize more semantic information.Based on the shortcomings of the aspect-level sentiment analysis task and the need to integrate external knowledge,the main research of this paper is as follows.(1)A multi-task learning model for aspect and opinion pair extraction based on knowledge enhancement.A knowledge-enhanced multi-task learning model for aspect and opinion pair extraction is proposed in order to correctly extract aspect and opinion items and identify the correspondence between them.Firstly,a pre-trained language model is used to generate semantically informative word vectors for the text.To achieve the knowledge enhancement effect,the semantic information of the knowledge graph is incorporated into the word vectors using masked attention,then the aspect and opinion items are extracted using a sequence annotation method based on distance attention and conditional random fields,and finally the extracted aspect and opinion items are matched in pairs to predict the correspondence.In order to strengthen the connection between the aspect and opinion item extraction module and the matching module,the model is jointly trained using a shared coding layer.In addition,the matching module uses real labels as input in the training process,and the results of the extraction module are used as input in the testing process.(2)Knowledge augmentation and prompt-based aspect sentiment classification model.To alleviate the discrepancy between the pre-trained language model and the aspect-level sentiment analysis target,prompt was introduced into aspect-level sentiment analysis by using pseudo-tags plus aspect words and opinion words to create prompt continuous templates,and splicing the prompt continuous templates with the source text as model input,in addition to using prompt-encoder to train the pseudo-tags to have semantic information.Then external knowledge about aspect words and opinion words is fused using a topic graph attention mechanism,and candidate tag words composed of sentiuent lexicon are predicted based on the hidden vector of fused external knowledge;finally the probabilities of candidate tag words are mapped onto the sentiment polarity distribution space using summed confidence scores.(3)Develop an online aspect-level sentiment analysis platform.In order to verify the validity of the above model,an online platform was developed to implement aspectlevel sentiment analysis and data statistics on the input text.Users of the platform are divided into ordinary users and administrator users.Ordinary users first enter the platform through the registration and login functions,then they can input text into the aspect-level sentiment analysis module for aspectual opinion extraction and aspectual sentiment classification through text input and file upload,and finally they can view the analysis results in their personal centre.In addition,administrator users can use the system monitoring module to achieve user management,model usage management and analysis result management.In summary,this paper addresses the current shortcomings in building deep learning models for aspect extraction and aspect sentiment classification respectively and builds an online user platform based on them.