Study On Aspect-level Context Sentiment Classification

With the development of Internet and social media in recent years,the number of network context data explosive growth.Studying the sentiment classification of these data is of great significance to the public opinion monitoring of government departments,product research of enterprises and product recommendation of merchants.The early context sentiment classification methods are mostly based on sentence level or even the entire chapter level,which is not suitable for scenes that require fine-grained analysis of context,resulting in context sentiment classification based on aspect-level.At present,the study object of most sentiment classification methods based on aspect-level is aspect-term,which is not a good way to achieve more abstract and higher-level sentiment analysis.These methods are very strict to the datasets,and need to annotate the aspects of the context and the specific location of them at the same time.So that,the usage scenarios of aspect-level sentiment classification is considerably limited by above methods.In response to the above problems,this thesis proposes an aspect-level context sentiment classification system.This system takes aspect-category as the study object and has very loose requirements for datasets.It only needs to provide the context and the set of aspects that the context may contain to get the sentiment polarity under each actual inclusion aspect.The specific work done in this thesis is as follows:(1)Introduces the principle of ABSA(Aspect Based Sentiment Analysis)in the natural language processing and some existing methods for the problem of context sentiment classification,including sentiment dictionary based methods,machine learning based methods and deep learning based methods,laying the foundation for follow-up work.(2)Extract and recognize the aspects contained in the context,and model that process as a multi-label classification task.In order to fulfill this task,this thesis proposes a label parameter based model called LP-LSTM(Label Parameter based LSTM),using the idea of deep learning.Experiments on benchmark dataset and the comparison to the state-of-art methods validate the improvement of accuracy by introducing the label parameter.The final result of the experimental results on the dataset reached 0.938,the recall rate reached 0.936,and the F₁ score reached 0.937.(3)Aspect-oriented sentiment classification.For the aspects that have been extracted and recognized above,we need to analyze the sentiment polarity for each aspect.This thesis finally achieve the sentiment classification of one or more aspects contained in a sentence by introducing an attention mechanism based on the LSTM.It is verified that the attention mechanism can effectively use the relationship between aspect and context by comparing with LSTM,Bi-LSTM and Bi-LSTM+aspect embedding models,and finally improve the accuracy of sentiment classification.The accuracy on the dataset reached 79.8%,and the macro?F₁ score reached 0.838.