| With the rapid development of social networks,users post comments with subjective emotions on major e-commerce platforms.Aiming at the massive user comment data,effectively mining the emotional information is of great value.Aspect-based sentiment analysis(ABSA)can analyze the sentiments and opinions expressed by users about different entities or attributes in the text.In recent years,deep learning has become the mainstream method for processing sentiment analysis tasks with its powerful semantic feature learning and extraction capabilities.However,at this stage,aspect-based sentiment analysis method based on deep learning still have shortcomings in many aspects.This thesis mainly focuses on aspect term extraction,aspect term sentiment classification and target-oriented opinion words extraction tasks in the ABSA field.Through some ABSA algorithms proposed currently,the problems and deficiencies in these methods are improved.The specific work and innovations are as follows:Firstly,regarding aspect term extraction,in view of the problem that a single word vector cannot obtain deeper semantic information in a specific field and CNN cannot consider the relationship between distant words,this thesis proposes a target extraction algorithm based on dual word embedding and self-attention coding.In this thesis,general embedding and domain embedding are used to enrich semantic information,self-attention coding mechanism is used to learn the internal structure of sentences.Experimental results show that double word embedding can enrich semantic information and increase F1 value;self-attention coding can take into account the semantic features of distant words.Secondly,regarding aspect term sentiment classification,in view of the problems that the previous neural network using recurrent neural network combined with attention mechanism brings about many training parameters and lacks interpretation of related syntax constraints and long-distance word dependence mechanism,this thesis proposes self-attention gated graph convolutional network.A multi-head self-attention mechanism is used to encode contextual words and targets.A graph convolutional network is established on the sentence’s dependency tree to obtain syntactic information and word dependencies.At the same time,this thesis also applies the pre-trained BERT language model to this task.Experimental results show that the multi-head self-attention mechanism can reduce network parameters and capture the semantic associations within sentences;graph convolution based on dependency trees can obtain sentence dependencies and improve the classification effect;BERT can further enhance the performance of the model.Finally,regarding the target-oriented opinion words extraction,considering the paired extraction of opinion words and target words and the previous model’s failure to pay attention to the positional relationship between them,this thesis proposes an opinion extraction algorithm based on aspect fusion and context focus mechanism.A CNN is used to extract features of the opinion words,then the features and sentence embedding are stitched into the network.At the same time,considering the feature that the position of the opinion word is often closer to the target word,the context focus mechanism is used to focus more on the aspect word and weaken the semantics of the distant word.Experimental results show that aspect fusion can strengthen the interaction between the target and the sentence,and the context focus mechanism can effectively improve the performance of opinion extraction.In this thesis,three different improved algorithms are proposed for three different subtasks in the ABSA field.Experimental results show that the performance of the algorithms proposed in this thesis have a certain degree of improvement compared with the existing algorithms for the same task.The research work in this thesis is of great value to the e-commerce platform to increase the diversity of product evaluation functions.It is helpful to improve the service and product quality of the e-commerce platform,provide users with personalized recommendations,intelligent search and other functions,and help users make the best decision making. |
