Research Of Aspect-based Sentiment Analysis For Review Text

With the rapid development of the Internet,social media platforms have become "gathering places" in people’s daily life.People can express their opinions or views about someone or somethings on social media platforms anytime and anywhere.For example,users will share their experience and evaluation of a certain product or service on Taobao,Jingdong and other shopping platforms.Citizens will express their views,suggestions or opinions on hot events or new policies issued by the government on Weibo,We Chat and other social platforms.These comments contain the subjective feelings of the publisher,so they are of great research value.Users can decide whether to consume or not based on the evaluation of others,enterprises can improve their products or services based on the evaluation of consumers,and governments can control the direction of public opinion through public opinions to maintain social stability.Therefore,aspect-based sentiment analysis,which can efficiently capture valuable information from massive comment texts,has become a hot research topic in the field of natural language processing.At present,most researches use deep learning models to solve the problem of aspectbased sentiment analysis,and have achieved remarkable results.However,there are still some problems.For example,some deep learning models,such as recurrent neural network and long short-term memory network,cannot learn globally dependent information due to their own structure limitations.Not taking full advantage of the text’s potential grammatical information.Underutilizing the advantages of multi-tasking frameworks,etc.This paper does some research on the method of aspect-based sentiment analysis based on deep learning.Enhancing the performance of the model in aspect-based sentiment analysis task by improving its task architecture,model or technology.The work includes:(1)Aiming at aspect term extraction task,we offer a joint extraction and detection method of aspect word and aspect category based on multi-task learning framework.In this method,the aspect word extraction task and aspect category detection task are placed under the multi-task learning framework,and the two tasks interact and promote each other through task sharing vector,thus enhancing the performance of the joint model in aspect word extraction task.In addition,the word co-occurrence information and syntactic dependence information are fully considered in this method to enhance the ability of modular line extraction of aspect words.(2)Aiming at aspect term sentiment analysis task,we design an aspect term sentiment analysis method based on target-dependent multi-head self-attention and word cooccurrence.In this method,an attention coding layer is designed to capture the semantic and grammatical information of the text from different levels.Then,a word co-occurrence map is constructed,and the word co-occurrence information is learned by graph convolutional neural network.Finally,a target-dependent multi-head self-attention is designed to further capture global features.For the two methods proposed above,this paper conducts comparative experiments on public data sets,and the experiment results prove the validity of two models.