Sentiment Analysis Of Review Text Based On Deep Learning

The existing text representation models are based on the relationship between words,which makes it difficult to capture the meanings of structured words or phrases in review texts.Deep sentiment analysis models usually treat the review text as a super-long sentence processing and ignore the hierarchical structure of the review.Because of the black box nature of the model,it is impossible to evaluate its ability to discriminate sentiment words or sentences in a concrete way.In reality,the small scale of review data sets often occurs,which makes the deep neural network difficult to apply,and the lack of professional knowledge of review text sentiment analysis makes it impossible to carry out the work of comment analysis.Currently,multi-task models are widely used in many fields,but most of the existing deep sentiment analysis models are single-task models.To solve the above problems,a Mark Recognition Strategy(MRS)is designed,which can introduce the structure information of review text into the deep learning model.Hierarchical Remark Neural Model(HMN)is also proposed for sentiment analysis.The model has a hierarchical structure of "word-sentence-review" which enables it to distinguish not only the sentimental level of words,but also the sentimental level of words.Sentences contribute to the whole review sentiment,and then the review text with complex structure is processed by introducing word-level and sentence-level markers,and the recognition ability of model HMN to sentiment words and sentences can be evaluated visually.For small review dataset scenarios,HMN-TF is proposed,and four transfer parameter tuning methods are proposed based on the hierarchical structure of the model.Sentiment analysis includes: sentiment classification,opinion summary extraction and other tasks.Based on HMN,a multi-task model HMN-MT is proposed,which uses opinion words to summarize the performance of abstract task promotion model in sentiment classification tasks.Experiments were conducted on the Amazon in Singapore and Test Freaks review datasets to evaluate the performance of MRS,HMN,HMN-TF,and HMN-MT,and to compare the leading-edge models of review text sentiment analysis with Accuracy and F1 Score indicators,HMN and mainstream sentiment analysis.The model is improved by about 1.6%,and the MRS can increase the model performance by 6.7%.HMN-TF and HMN-MT are also superior to the comparative frontier model.