Asnect-level Sentiment Analvsis And Research On Deep Neural Network

Aspect-level sentiment analysis,as a kind of fine-grained sentiment analysis,is widely studied nowadays.Its main tasks are:identifying aspect words in sentences and sentiment polarity analysis based on aspect words.In traditional sentiment analysis,the aim is to analyze the sentiment polarity contained in a paragraph(or sentence),but this is based on the assumption that only one sentiment polarity is contained in the entire text.However,in real scenarios,this is often not the case.For example,there is an online review about a restaurant:"The service of this restaurant is good,but the taste of the food is terrible",in this sentence,"service" and "food" contain two different emotional polarities.For restaurant service,the emotional polarity is positive,while for restaurant food,the emotional polarity is negative.From this point of view,traditional text-level sentiment analysis methods cannot meet the current sentiment analysis needs.Aiming at the above problems,this paper has carried out relevant exploration,the main research work is as follows:(1)An aspect-level sentiment analysis model based on parallel convolutional neural network is proposed.After the model maps each word to a low-dimensional vector space and obtains the word vector,it constructs a graph according to two strategies of dependency tree and emotional intensity,so that the graph volume The product neural network is able to learn various internal connections in the text.In addition,since the graph convolutional neural network inevitably introduces noise when aggregating node information,in order to filter the noise generated by the graph convolutional neural network,the model also introduces a gate mechanism that can filter out aspects irrelevant to the aspect words.noise.Finally,in order to fully capture the connection between different graph convolutional neural networks,the model utilizes three attention mechanisms to calculate feature vectors of different aspects,and then predicts the sentiment polarity according to the three feature vectors.(2)E2E-CO-SFTP is proposed.This model combines aspect word extraction and sentiment analysis into a sequence tagging task,and simultaneously predicts the position information and emotional tendency of aspect words during tagging..It uses the pre-training model BERT as the word vector model,maps the words into low-latitude space vectors,and then inputs the word vectors into the CO-SFTP layer,uses interactive attention to mine the internal relationship between words,and uses self-attention to capture the relationship between words and Intrinsic relationship between sentences,and finally integrate the two to increase the global vision of the model.Finally,the conditional random field is used to learn the sequential constraints,and the predicted labels are predicted by maximizing the conditional probability distribution.