Attribute-level Sentiment Analysis Research For Product Online Review Text

The popularity of e-commerce has led to online shopping becoming a major form of shopping in people’s daily lives,and consumers have accumulated a large number of review texts with sentiment preferences related to products in major e-commerce platforms,and the emotional information of the all-round mining of the comment text has important significance and reference value for manufacturers and consumers.Traditional sentence-level or document-level sentiment analysis cannot accurately reflect consumers’ sentiment towards a certain attribute of a product,and it is difficult to meet the refined needs of consumers and enterprises.In this thesis,we investigate the attribute-level sentiment analysis of online product review texts to uncover the product attributes that consumers care about and their corresponding sentiment polarity.In this thesis,we take mobile phone online review text as the research object,and carry out attribute-level sentiment analysis research in terms of both attribute extraction and attribute-specific sentiment classification.The main work of this thesis is as follows.:(1)Constructing the dataset.This thesis first uses web crawling technology to crawl a large number of online review texts of mobile phones from e-commerce platforms,and then pre-processes and visualises the data to analyse the characteristics and content of the research topic of this thesis.(2)Product attribute extraction.In order to extract potential product attributes from online review texts,this thesis uses a method based on syntactic relations to extract product attribute words.The method first uses dependent syntaxto analyse the syntactic structure of the text,and constructs an attribute word extraction template suitable for the review text of this thesis according to the summarised dependency relations and lexical collocation rules,so as to extract the set of attribute words.At the same time,in order to summarize the corresponding attribute topics from the set of attribute words with similar word meanings,this thesis proposes a affinity propagation clustering algorithm based on cosine similarity(APCCS)to cluster the attribute words and extract the corresponding cluster center as the attribute subject word of the corresponding class cluster.(3)Attribute-specific sentiment classification.In order to classify sentiment separately for several attribute topics contained in each review text,this thesis proposed a multi-headed self-attentive gated convolutional network model and makes comparison experiments with other benchmark models on the labeled dataset.The results show that the model proposed in this thesis is better than other models in both accuracy and Macro-F1 values.Finally,we analyse the strengths and weaknesses of the product in each of the fundamental attribute dimensions and make recommendations for targeted product iterations,using the product attribute theme as the basic dimension for the sentiment polarity of the online review text.