Research On Text Sentiment Analysis Based On Deep Learning

With the development of Internet technology,human society is entering a stage of highly intelligent and informatized.As a condensed information carrier,text has always been the main medium for people's information exchange.At present,a large amount of text data has been accumulated on the Internet,which contains people's emotional tendencies towards goods and services.The emotional orientation in the research text can help merchants or service providers make relevant decisions,and has important commercial value and academic research significance.Aiming at the problem of insufficient emotion word extraction ability in existing text sentiment analysis methods and the out-of-vocabulary(OOV)problem of pretraining word vectors,this paper proposes a SAT-Bi LSTM neural network model.which combines multi-head self-attention mechanism and character-level embedding technology.By using character-level embedding and pre-training word vectors,the input of the model can deal with spelling errors and rare words flexibly,thus effectively solving the OOV problem.A Twitter sentiment analysis method based on the combination of multi-head self-attention mechanism and two-way LSTM is proposed.In the sentiment analysis of Twitter short texts,a multi-head self-attention mechanism is introduced,and experiments have verified the advancement of the above model in sentiment analysis tasks.Experiment with this algorithm on data sets in different fields,and compare other algorithms.The results show that the classification accuracy of the SAT-Bi LSTM model is higher than that of other baseline models.Aspect sentiment analysis aims to extract aspect terms and predict the sentiment category of opinions.It includes two subtasks:aspect item extraction and aspect sentiment classification.However,previous research treats them as two independent tasks and solves them separately,which has limitations in practical applications.In this paper,combining the requirements of two subtasks,a span-based aspect-based sentiment analysis framework is proposed.This framework is a simple and effective joint model that can generate various aspects of the input sentence and the corresponding sentiment polarity.The two-door control unit used to extract the corresponding representation of each task can better process the sequence information,as well as the interaction layer used to consider the relationship between the representations.Experiments on three data sets show that the proposed framework outperforms the baseline model.Past work to improve document-level sentiment analysis by encoding user and product information has been limited to considering only the current text.When other available review texts are included in emotion prediction,these texts may have guiding significance for emotion prediction.First,merge all available historical reviews of related review authors;second,the survey contains historical reviews associated with the current product;finally,it explicitly stores the representation of reviews written by the same user and the same product,and forces the model to remember All reviews of a particular user and product.In addition,the layered architecture used in previous work has been deleted so that the words in the text can directly focus on each other.Experiments on IMDB,Yelp 2013 and Yelp 2014 datasets prove the effectiveness of the model.