Aspect-level Sentiment Classification Approach Based On Deep Learning

With the widespread attention of opinion mining and sentiment analysis in academia and industry,and the increasing popularity of e-commerce such as Jingdong and Taobao,a large number of user comment texts have been generated,which has brought valuable resources for scientific research.These massive review texts contain the opinions expressed by users on goods and services.These review information can become valuable resources for producers to improve the quality of products and services,and have great commercial value.However,the exponential growth of review data has made it impossible for humans to obtain useful information from massive amounts of data.Existing studies have shown that sentiment analysis tasks based on deep learning can effectively extract a given aspect of sentiment from the text,thereby extracting the information needed by the user.This paper mainly studies the aspectlevel sentiment analysis task in sentiment analysis tasks.It is a fine-grained sentiment analysis task,which aims to provide more detailed user feedback information than traditional sentiment analysis.According to the application scenarios and the characteristics of comment data,the main research contents and contributions are as follows:1、Aiming at the problem of sentiment prediction errors caused by inaccurate aspect feature extraction in aspect-level sentiment analysis,we introduce the sentiment category features.At the same time,aspect-guided sentiment extraction and sentiment-guided aspect extraction are used learn the potential connection between aspect and sentiment.This paper propose the aspect sentiment analysis method based on collaborative extraction hierarchical attention network.When the aspect attention layer cannot accurately extract the uncommon aspect features,it can accurately identify the sentiment expressed by the user according to the given aspect sentiment category.Experiments have proved that this collaborative extraction method can better learn the potential connection between aspect features and sentiment features,and can reduce sentiment misjudgments caused by the extraction error of aspect features.2、The introduction of the attention mechanism will make each word in the review sentence more or less assign some attention weight,but only a few words are related to the task,which is the inherent noise of the attention mechanism.Therefore,this paper introduces a sparse regularization mechanism on the basis of the hierarchical attention network in 1,to make the distribution of attention weights more sparse,so that the distribution of attention weights is concentrated on a few task-related words.So as to alleviate the inherent noise of the attention mechanism to a certain extent.In addition,this paper further expands multi-task learning,using aspect detection tasks as auxiliary tasks to optimize the loss function,speeding up the backpropagation process and improving model performance.The experimental results prove that the multi-aspect sentiment analysis method based on the constrained hierarchical attention network proposed in this paper achieves better results.3、From the data point of view,it is found that in the review data,a user review often contains opinions on multiple aspects,and the sentiment of multiple aspects are connected by some conjunctions,and the sentiment polarity between them is dependent.Moreover,the aspect terms of different aspects are irrelevant.Therefore,this paper introduces orthogonal regularization on the basis of 2 to limit the distribution of attention weights,so that the attention weights of different aspects concentrates on different parts of the sentence,so that aspect features can be extracted more accurately.For the sentiment dependence,this paper inputs the different aspects of sentiment features extracted from the sentence into a Bi LSTM in order to learn their sentiment dependence.This is an aspect-level sentiment analysis method based on the aspect-aware hierarchical attention network proposed in this paper.The results on the Sem Eval data set also prove the effectiveness and reliability of this method.In summary,this paper uses deep learning methods to focus on the aspect-level sentiment analysis techniques based on collaborative extraction hierarchical attention networks,constrained hierarchical attention networks,and aspect-aware hierarchical attention networks.These technologies and research results can better help us to accurately and quickly identify the aspects mentioned and the sentiment expressed by users from the massive review texts,and play a vital role in product updates and recommendations.