Research On Sentiment Classification Of Social Comment Short Text

Text sentiment classification involves many research fields such as natural language processing and deep learning,and plays an important role in many fields,especially in the service industry(such as the hotel industry and the airline industry).By sentiment classifycation of customers' social media comments,companies can conveniently acquire customers' sentiment tendencies and timely adjust service strategies,which is conducive to provid better services for customers.The hotel field,and the aviation field are taken as the research object in this thesis,and the aspects of short text sentiment classification model based on word embedding,domain sentiment lexicon construction,text sentiment value calculation be researched and explored.The main research contents are as follows:(1)For the problem of token masking strategies of BERT lead to insufficient feature extraction,Chinese short text classification model based on ERNIE is proposed in order to optimize Chinese short text classification algorithm based on BERT.Firstly,text preprocessing,secondly,get the text semantic representation by using ERNIE,finally,put the semantic feature vector into Softmax classifier to finish sentiment classification.In the experimental process,hyperparameters of classification model were adjusted repeatedly,and many different combinations of hyperparameters were tried.The contrast experimental results show that the proposed Chinese short text classification model based on ERNIE is superior to the control group,which proves the effectiveness of the proposed method.(2)The construction method of domain sentiment dictionary based on seed word expansion has a low ability to extract the semantic information of word context,which leads to the poor domain property of the constructed domain sentiment lexicon,therefore,an ERNIE-based construction method of domain sentiment lexicon(ERNIE-Senti L)is proposed.In ERNIE-Senti L,the construction of a domain sentiment lexicon will be considered as a sentiment classification task for a word or phrase.Firstly,ERNIE was used to extract candidate sentiment words' semantic feature representation that input into Softmax classifier,Softmax can output each candidate words' confidence(probability)of the category(positive and negative),finally,the candidate words were classified according to the confidence,and the domain sentiment lexicon was constructed.The contrast experiments show that the domain sentiment lexicon,by using ERNIE-Senti L,has good domain expressiveness and sentiment recognition effectiveness.(3)The traditional method of sentiment value calculation has the problem of insufficient feature word extraction,therefore,a text's sentiment value calculating method by combining modal is proposed.Firstly,a modal adverbs lexicon is constructed,which includes weakening,strengthening,and negative modal adverbs,then the modal adverbs are incorporated into the text's sentiment value calculation rules.The contrast experiments show sentiment classifycation model constructed by using the sentiment value calculation method combining modal adverbs is superior to that constructed by using the traditional sentiment value calculation method.