Research On Cross-lingual Text Sentiment Classification Based On Deep Learning

With the rapid development of Internet technology,people are more and more keen to express their views on an event or thing on the Internet.There is huge commercial value behind these comments.Therefore,in recent years,the study of text sentiment analysis has attracted more and more attention.However,some languages start later than other languages,lack of high-quality corpus resources,and manual labeling requires huge human and material resources,which hinders the research of text sentiment classification technology to a certain extent.The cross-lingual text sentiment classification task is to use a language rich in corpus resources to assist another language lacking in corpus resources to achieve text sentiment classification.In order to further improve the performance of cross-lingual text sentiment classification,this paper integrates and improves various aspects,and the following cross-lingual text sentiment classification methods are proposed:(1)Aiming at the problem the traditional vector representation of single language words can't perform interactive learning well between the two languages,a crosslingual sentiment classification method based on adversarial long short term memory network is proposed.This method sets up independent feature extraction networks of source language and target language and bilingual shared feature extraction network,establishes the relationship between source language and target language,and reduces the semantic gap between the two languages.At the same time,using the adversarial idea,a language classifier is set up in the shared feature extraction network to make it indistinguishable whether the feature is from the source language or the target language,in order to obtain invariant bilingual features,so as to achieve better knowledge transfer between the two languages.Compared with previous research methods,this method not only retains the independent features of bilingualism,but also obtains the invariant features of bilingualism.Experiments are conducted on the NLPCC 2013 cross-lingual sentiment classification dataset,the result shows that this method improves the performance of sentiment classification.(2)Considering that the sentiment lexicon is still an indispensable resource,and the contextual information of sentiment words contributes greatly to the sentiment polarity of the whole corpus,a cross-lingual text sentiment classification method that combines local and global features is proposed.This method combines the sentiment lexicon,the convolutional neural network is used to obtain the contextual features of the sentiment words as the local features of the whole corpus.At the same time,the bidirectional long short memory network with attention mechanism is used to obtain the global feature of the whole corpus.Finally,the local feature and the global feature are merged as the final classification feature,which are input to the classifier for text sentiment polarity classification.Experiments are conducted on datasets containing multiple languages,and the results show this method improves the performance of sentiment classification.