Research On Deep Learning Methods For Fine-grained Sentiment Analysis Of Chinese Online Reviews

Sentiment analysis is a natural language processing task of mining emotional information from text data.The early coarse-grained sentiment analysis method is mainly used to judge the emotional polarity of the article or sentence as a whole,such as negative,neutral,positive,and has a good application in the fields of stock evaluation analysis and public opinion analysis.However,in the online commentary texts of online shopping and service appointments,the overall emotional polarity is unable to accurately capture the specific advantages and disadvantages of the goods or services,and ultimately does not fully exploit the emotional information of the text.Therefore,many merchants and users on the Internet platform have put forward new demands on the sentiment analysis task,that is,they hope to further obtain the sentiment analysis results corresponding to a certain aspect of the specific review target,such as the "taste" aspect of "dish" or "price".This fine-grained sentiment analysis method is also called aspect-level sentiment analysis because it ultimately gives an aspect-based sentiment analysis result.This paper will focus on the Chinese online comment text with a length of more than 200 chars,and develop a deep learning method for fine-grained sentiment analysis.The main research work of this paper is as follows:(1)In the problem of local feature separation and extraction of text,this paper proposes a word selection mechanism(WSM).Since the words used to express the emotional information corresponding to a certain aspect account for a relatively small amount in the whole text,how to extract keywords from the text will be the focus of the aspectlevel sentiment analysis.The word selection mechanism proposed in this paper can effectively extract words related to sentiment analysis tasks from lengthy texts.Macro F1 value is adopted as the evaluation index in the experiment,and the average results in 20 aspects of the text show that CNN+WSM increased by 1.2% compared with CNN,and LSTM+WSM increased by 2.6% compared with LSTM.(2)In the sampling method of unbalanced dataset,this paper proposes a sampling algorithm based on exponential compression.When sampling unbalanced data,the algorithm can control the sampling proportion effectively,so that the sampled data can not only maintain the characteristics of the original unbalanced distribution to some extent,but also be relatively balanced.The optimal compression coefficient was found through lattice search,and finally macro F1 value of the model was improved.The experimental results show that the sampling algorithm proposed in this paper can effectively improve the under-fitting of a few classes of training models when training unbalanced data.(3)On the research of multi-aspect sentiment analysis,this paper proposes a model that can simultaneously analyze multiple aspects of the text,namely,aspect-oriented multi-label learning(AOML)sentiment analysis language model.Different from the existing mainstream sentiment analysis models ATAE,GCAE,etc.,this model does not need to add additional aspect information,but can actively find and locate aspect information from the text,and then conduct emotional analysis under the guidance of aspect information.The results show that the macro F1 value of AOML+WSM model proposed in this paper is 2.9% higher than GCAE and 0.9% higher than ATAE.