Aspect-level Sentiment Analysis Research Based On Deep Learning

With the rapid development of the Internet age,netizens can post their own unique comments on various commodities on different life websites,and they can also put forward their own opinions on a social phenomenon in social networks.The text comment information published by users contains many sentences with sentiment inclinations.After sorting and researching,these sentences have high research value in related fields such as commodities and society.Therefore,text sentiment analysis,as a sub-field of natural language processing research,has received extensive attention in the past few years.Aspect-level sentiment classification is a fine-grained work in sentiment analysis.Compared with traditional sentiment analysis,the goal of aspectlevel sentiment analysis is to predict the sentiment categories of different target sentiment words in the text,so the level of text content mining deeper.Sentiment analysis models based on deep learning have achieved many breakthrough results in text feature extraction,and have become the mainstream research direction in the field of sentiment analysis.In aspect-level sentiment analysis tasks,the memory network model has achieved excellent performance,and the introduction of the attention mechanism has further improved the performance of sentiment classification.However,there is still room for improvement in aspect-level sentiment analysis: when the text is too long,it is difficult to capture the long-distance sentiment features accurately;the attention mechanism lacks attention to the target aspect word position at the same time,ignoring part of the sentiment feature information extract.Therefore,this paper designs an interactive attention mechanism based on the memory network,and conducts research and evaluation on the interactive attention mechanism in the deep memory network and long-short-term memory network models.The main research contents of this paper are as follows.(1)Aiming at the problems that the sentiment characteristics of sentiment analysis texts are too long and long-distance are difficult to capture and the attention mechanism lacks attention to the word position of the target,this paper proposes a deep memory network-based cyclic interactive attention sentiment analysis model(Mem-IAM).Mem-IAM uses Bi-LSTM neural network as a memory network unit to capture the feature information of long-distance texts,and designs a text weight function to emphasize the connection between target aspect words and context words.And designed a circular interactive attention mechanism to focus on the target keywords and contextual information.Mem-IAM is composed of multiple computing layers that share parameters.The interactive attention layer is introduced to process the complex features in emotional sentences,and the important information between context and aspect words is continuously transferred to the next layer,thereby increasing the emotional words in the context The weight ratio of,can realize emotion classification more accurately.(2)The interactive attention mechanism has achieved considerable results in sentiment analysis.Therefore,in order to further improve the application of the interactive attention mechanism in deep learning,an interactive attention model based on Bi-LSTM and position-aware mechanism(Bi-LSTM-Pos-IAM)and sentiment analysis model based on Bert and Bi-LSTM interactive attention mechanism(Bert-BiLSTM-IAM).Our model introduces the Bi-LSTM neural network,which can read the relevant information of the context and aspect words from two directions,and model the context and aspect words respectively.In addition,a position-aware mechanism is added to the context embedding vector to capture the relationship between contexts.The location index perception mechanism can use the location information between aspect words and other words to determine terms related to sentiment polarity.At the level of attention mechanism,the main idea is to extract the features of words and sentences and learn alternately to obtain a more reasonable attention weight distribution.In addition,in order to make better use of aspect information,we embed the input aspect into each word input vector as a supplement to improve the accuracy of the proposed model.The experimental results show that compared with the baseline model,the accuracy and F-measure of our algorithm perform better.