Research On Fine-grained Sentiment Analysis And Multi-type Sarcam Detection Of Web Comments

Web comments contain the user's sentiment information about service and product.It plays an essential guiding role in business operations.Therefore,sentiment analysis of web comments has fundamental practical application value.For fine-grained analysis of users' sentiment polarities about certain aspects in web comments,aspect-based sentiment analysis(ABSA)is a very effective method.However,the current research method has the following drawbacks.More and more complicated web comments cause the sentiment features extracted by the recurrent neural network(RNN)to become unreliable.Because RNN may forget some critical sentiment information,at the same time,the aspect characteristics obtained by the current average aspect vector method may be wrong.Besides,more and more Internet users like to use sarcasm to express emotions,which makes it more challenging to analyze sentiment in web comments.However,the existing sarcasm detection method is too one-sided.It can only detect sarcasm of web comments,or detect in combination with the user's background information.When the emotional change in web comments is relatively significant,the context information added to web comments will lead to wrong results;when web comments information is insufficient,it is impossible to detect whether the user used sarcasm rhetoric.Therefore,this paper conducts in-depth research on aspect-based sentiment analysis and sarcasm detection.The main work and research results of this paper are as follows:(1)Proposed a memory network model based on the Transformer(TF-MN).In TF-MN,to correct the error caused by the average aspect vector,the sentiment analysis task is transformed into a question and answer process,which includes four modules:context,question,memory and answer module.In the context and question module,this paper uses a Transformer based on a self-attention mechanism to capture the global sentiment features.Also,words such as partial nouns,prepositions,and adverbs will affect the extraction of emotional features of web comments.Therefore,in the memory module,this paper uses a memory network model based on the local attention mechanism to filter irrelevant sentiment features.In particular,to shield the influence of foreign words,this paper improves the memory network to optimize the extraction of sentiment features of web comments.We tested the TF-MN model on the Chinese Weibo and Semeval-2014 datasets,and the results verified the effectiveness of the model.(2)Proposed a sarcasm detection method based on multi-level semantic capsules.This method comprehensively considers the sarcasm based on web comments(i.e.,verbal sarcasm)and the sarcasm combined with the context of web comments(i.e.,situational sarcasm).First,SenticNet is used to offset the sentiment weight of web comments to obtain semantic capsules to solve verbal sarcasm.Then,through scaled dot product attention mechanism,web comments and contextual information were fused to obtain a semantic capsule for situational sarcasm.Next,the two semantic capsules are combined to form a low-level semantic capsule layer.After that,this model obtains a high-level semantic capsule layer through EM routing.Finally,use the SoftMax function to output the results of sarcasm detection.The model proposed in this paper has obtained significant experimental results on three datasets,proving the superiority of the model.(3)Developed a sentiment analysis system for Qiandaohu tourism web comments.First,develop a crawler management system capable of crawling dynamic webpage data with Selenium as the crawler core,including adding,deleting,crawling,storing,query,counting and checking functions.The crawled data mainly comes from sights and hotel plate of Ctrip,and restaurants plate of DaZhongDianPing.Besides,by tagging a small amount of web comments data as a training dataset,a pre-trained model is created,and other web comments are fine-grained emotionally scored,enriching the sentiment information of web comments.