Fine-Grained Sentiment Analysis And Application Based On Text Themes

Subject-based text sentiment analysis technology has gradually become a significant topic,the main task is to explore the subjects contained in user comments and the emotional preferences of these subjects.The sentiment in this article consists of three levels:satisfaction,neutrality,and dissatisfaction.Different from traditional sentiment analysis and aspect-level sentiment analysis,"fine-grained" is reflected where there are multiple<subject words,sentiment words,sentiment analysis>triples as a result;In addition,the subject is no longer an "aspect" level,but an arbitrary word which is more random.Based on real Internet commodity review data and model fusion,this paper proposes a set of algorithm fr-ameworks including entity extraction,subject matching sentiment and subject sentiment analysis which has achieved stable and significant result using variety metrics under cross-validation conditions.The main work of this paper is as follows:(1)In subj ect word and sentiment word extraction,we have compared various machine learning algorithms and deep learning algorithms and made Bi-LSTM combined with Conditional Random Field as the final algorithm structure;referring to the idea of word embedding,dictionary word embedding were proposed in this paper which has resulted in significant improvements in recall and accuracy.(2)We proposed hierarchical matching strategy which containing two stages:in the R(Recall)stage,all the<subject words,sentiment words>are first roughly selected to improve the recall rate;in the P(Precision)stage,in order to improve the accuracy rate,we use deep neural network based on current pairs to select in detail.Experimental results verify the role of the strategy.(3)As for sentiment analysis,considering the difference of context,we have made sentiment words,contextual word vectors and other related information as feature.A new algorithm framework is proposed and achieved significant experimental results.