The Research Of Fine-grained Sentiment Analysis Of User Reviews Based On Neural Network

With the development of the information age,an amount of user reviews has emerged on platforms such as e-commerce.This information conveys the users' views and sentiments,which is of great significance to the decision-making of consumers and businesses.Sentiment analysis is to sort out,summarize,and analyze the sentiments expressed by users.Compared with judging the sentiment tendency of the whole sentence,fine-grained sentiment analysis is based on each aspect which is involved in the sentence as the main body,and judges the sentiment tendency based on a given aspect.Fine-grained sentiment analysis includes two sub-tasks: aspect term extraction and target?dependent sentiment classification.In order to capture sentiment information from a large number of comments,automated text mining methods need to be researched to reduce the time for manually organizing and analyzing data.With the applications of methods based on neural network in natural language processing,it is found that this method can automatically capture valid text features,which solves the problem that traditional methods rely on grammar rules and feature engineering.This paper proposes models based on neural network for the two sub-tasks of fine-grained sentiment analysis:(1)The aspect term extraction aims to label the aspect term in the user reviews.This task can be regarded as a sequence labeling task,and the output sequence is labeled as a B-I-O mode.The labeling mode has certain regular constraints,such as label I will not appear directly after label O.Aiming at the problem that the constraints of annotations are not taken into account in the model in the existing methods,this paper proposes a model based on historical sequence attention mechanism.First,the general embedding and the field embedding are used to construct the input matrix.Then the model uses multi-layer convolutional neural networks to model review sentences.After that,the historical sequence attention mechanism is used to fuse the important features in the previous article with the current features.In this way,the current output label will be affected by the previous labels.Finally,the output features are fed into the classifier to realize the labeling of the output sequence.Experiment shows that the model has improved the F1 index on several datasets.(2)Target-dependent sentiment classification aims at identifying the sentiment polarities of targets in a given sentence.A common model structure is the recurrent neural network with the attention mechanism.Considering that the aspect is often close to the opinion word,the recurrent neural network will introduce the long-distance information into the current feature.And it is incapable either of modeling complex contexts or of processing data parallelly.To solve the above problems,a model based on the target mechanism and multi-layer convolutional neural network is proposed.First the model uses the convolutional neural network to process contexts in parallel,and uses a multi-layer structure to model the context multiple times.Then the neural network is able to explicitly learn the sentiment intermediate representation relative to a particular aspect via an attention mechanism.Eventually,we integrate these features to form a final sentiment representation,which will be fed into the classifier.Experiment shows that the model has improved the accuracy on several datasets.Finally,fine-grained sentiment analysis is implemented using the above two models.Experiments show that the F1 index of the model on the public dataset has improved.