基于ALBERT-HACNN-TUP模型的文本情感分析  

Text Sentiment Analysis Based on ALBERT-HACNN-TUP Model

作　　者：高玮军[1] 赵华洋 李磊 朱婧 GAO Wei-jun;ZHAO Hua-yang;LI Lei;ZHU Jing(College of Computer and Communication,Lanzhou University of Technology,Lanzhou Gansu 730050,China)

机构地区：[1]兰州理工大学计算机与通信学院,甘肃兰州730050

出　　处：《计算机仿真》2023年第5期491-496,共6页Computer Simulation

摘　　要：随着网络中多元化、碎片化的文本数量增多,传统模型对此类文本进行情感分析时,存在长距离语义信息挖掘不够充分、深层次的特征提取不够完整的问题。为解决上述问题,提出了基于ALBERT-HACNN-TUP(A self-supervised learning model based on a Lite BERT and text universal pooling a hierarchical attention convolutional neural network)的情感分析模型。模型首先使用ALBERT预训练语言模型提取更长距离的语义信息;其次改进CNN的卷积层,提出了一种分层注意力卷积神经网络(HACNN),根据卷积层提取特征信息的重要程度进行动态权重调整,进一步突出文本的情感极性词;再利用池化层Text Universal Pooling(TUP)动态学习池化权重,对不同通道进行提取和融合,最大程度保留了文本更深层次的情感特征,尤其对含有复杂语义的反讽文本有更好的效果。在不同数据集上进行了实验。仿真结果表明,上述模型提高了运行效率,具有良好的泛化性与精确度。With the increase in the number of diversified and fragmented texts in the Internet,when traditional models perform sentiment analysis on such texts,there are problems that long-distance semantic information mining is not sufficient and deep-level feature extraction is not complete.In this regard,a sentiment analysis model based on ALBERT-HACNN-TUP(A self-supervised learning model based on a Lite BERT and text universal pooling a hierarchical attention convolutional neural network)is proposed.The model first uses the ALBERT pre-training language model to extract longer-distance semantic information;secondly,it improves the convolutional layer of CNN and proposes a hierarchical attention convolutional neural network(HACNN),which extracts feature information according to the importance of the convolutional layer Perform dynamic weight adjustment to further highlight the emotional polar words of the text;and then by utilizing the pooling layer Text Universal pooling(TUP)to dynamically learn weights,different channel texts are extracted and fused,maximizing the preservation of deeper emotional features in the text,especially for ironic texts with complex semantics.Experiments were conducted on different data sets,and simulation results show that the model has improved operating efficiency,and has good generalization and accuracy.

关 键 词：情感分析 深度学习 预训练语言模型 卷积神经网络 反讽句式 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]