Research On Aspect-Level Sentiment Classification Method For User Reviews

In recent years,due to the transformation and upgrading of people’s consumption demand,the domestic Internet service industry has developed rapidly,accumulating a large amount of user comment information.Sentiment analysis on these comments will help the machine fully understand the attitudes,opinions and emotional tendencies contained in the comments,and help the network platform to understand user needs,social hotspots,public opinion trends,etc.,which is of great research significance.However,traditional methods can only infer the sentiment polarity of a document or sentence as a whole,and cannot analyze specific aspects of the sentence.Aspect-level sentiment classification solves this problem,it can infer more comprehensive and deeper sentiment polarity in text sequences,and can better reflect the real thoughts of users.Existing researches on aspect-level sentiment classification ignore the sentiment information and syntactic structure information of reviews under specific aspect categories,and fail to fully learn the semantic,syntactic,and sentimental relationships between words.This thesis solves this problem and proposes aspect-level sentiment classification methods based on pre-trained sentiment embedding and dual-graph attention network,respectively.The main work includes:(1)A deep learning method based on pre-trained sentiment embeddings is proposed.It first uses an adversarial learning method for training sentiment word embeddings and a two-layer bidirectional long short-term memory network to extract sentiment embeddings of the entire sentence under specific aspect categories.This vector is then dynamically combined with the semantic information from the last layer of the BERT model.Finally,in order to improve the performance of the model,linear weighting and multi-head self-attention mechanism are used to guide the model to pay more attention to the sentiment and semantic information of specific aspect categories.Multiple sets of experiments are conducted on the Chinese patient review dataset,and the results verify that the model outperforms other state-of-the-art methods.(2)Considering the sequence context information in reviews and syntactic dependencies between words,an aspect-level sentiment classification method based on bi-graph attention network is proposed.The method first constructs two graph networks to describe the sequence context information and syntactic structure information of sentences,then utilizes an attention mechanism to aggregate information from neighborhood nodes within a single graph,and uses a bi-affine module to coordinate heterogeneous information between sequence graph and syntactic graph.Finally,this thesis utilizes the aspect-specific mask and retrieval-based attention mechanism to make the model pay more attention to the sentiment information of specific aspect categories,thereby inferring the sentiment polarity of the specific aspect category.Experimental results on four datasets published in the SemEval 2015 and 2016 competitions show that the performance of this model is significantly better than other state-of-the-art models in aspect-level sentiment classification.In this thesis,aspect-level sentiment classification of Chinese patient reviews and English product reviews can mine more fine-grained information in sentences,which has broad application prospects.It not only provides users with rich decision-making information,but also reminds the platform to improve the defects existing in a certain aspect,which promotes the improvement of Internet service quality in China.In response to this research,we will continue to explore in the direction of reducing error propagation,avoiding over-parameterization,and multimodal learning in the future.