融合Stacking和深度学习的中文产品评论情感分析

Emotional Analysis of Chinese Product Reviews Based on Stacking and Deep Learning

大量涌现的电商产品评论对企业制定商业决策十分有利,BERT应用在英语文本情感分析中取得了不错的效果。针对中文电商产品文本评论提出了一个新的融合Stacking集成思想和深度学习算法模型。首先在文本信息特征提取层使用Chinese-BERT-wwm生成含有丰富语义信息的动态句子表征向量,Chinese-BERT-wwm是专门针对中文特点改进后的预训练模型,具有稳健的中文文本特征信息提取能力,其次该层同时设计了TextCNN和BiLSTM捕获文本中局部关键信息特征与语序信息特征,并将这些特征拼接在一起以获得更全面丰富的句子信息,最后基于Stacking集成学习思想使用SVM对该特征进行分类。为了评估模型效果,人工标注3万条具有三类情感极性的中文电商产品文本数据进行实验,该数据集可广泛用于中文情感分析领域。实验结果表明,与基线模型相比,提出的模型可以有效提高中文文本情感极性分类任务的准确率。

With the rapid development of e-commerce platforms,online shopping has become the mainstream consumption mode of users.Thus,as for enterprises,how to use a large number of emerging e-commerce product reviews has become the key to making business decisions.In addition,the application of BERT in English sentiment analysis has achieved excellent performance.Aiming at improving the analysis effect of Chinese e-commerce product text review,this paper proposes a new algorithm model combining Stacking integration idea and deep learning.First,in the text information feature extraction layer,Chinese-BERT-wwm is utilized to generate dynamic sentence representation vectors with rich semantic information,which improves robustness of Chinese text feature information extraction.Then,TextCNN and BiLSTM are designed in the same layer to capture features of local key information and word order information in texts.In addition,these features are stitched together to get complete sentence information.Finally,SVM is utilized to classify features based on Stacking integrated learning idea.The analysis and verification of the model are conducted based on 30000 text data of Chinese e-commerce products with three types of emotional polarity,which can be widely used to analyze Chinese sentiment.The results show that comparing with the baseline model,the proposed model can effectively improve the accuracy of Chinese text emotional polarity classification tasks.