Application of the Deep Pretrained Language Model Processing Method in Social Network Sentiment Analysis
Application of the Deep Pretrained Language Model Processing Method in Social Network Sentiment Analysis
复制标题
DOI：
10.1155/2022/2469693
复制
发表时间：
2022-08
影响因子：
--
通讯作者：
A. Yana
中科院分区：
文献类型：
--
作者：
A. Yana
在社交网络中，用户可以管理自己的社交网络和社交身份，发布各种主题的信息，并通过好友关系获取其他用户发布的信息。由此产生的大量文本数据吸引了越来越多的学者对其进行研究，文本情感分析已经成为社会网络数据分析的热点，在学术领域、社会领域、商业领域都有重要的应用价值。基于预训练的思想，对BERT（Bidirectional Encoder Representation）模型中深度预训练任务的随机词掩蔽算法进行改进，以提高模型预训练的效率和稳定性。其次，设计了一个新的句子判断预训练任务，使模型能够衡量句子的流畅程度，从而使BERT模型能够更好地理解上下文的语义。借鉴注意力机制的思想，构建了一种在门控卷积中加入注意力权重的深度学习框架，并采用特殊的注意力权重方法增强语义信息。其次，结合门控卷积和注意力机制对方面相关语义信息和文本完整语义信息进行建模。最后，对社交网络的情感分类器层进行分类，使用Softmax函数完成消极、积极和中性的多重分类，并计算情感分类结果。通过将优化后的卷积神经网络循环优化网络应用于实际的单任务和多任务，验证了优化后的卷积神经网络和循环神经网络应用于社交网络情感分析的可行性。
In social network, users can manage their social network and social identity, publish information on various topics, and obtain information published by other users through friend relationship. The resulting large amount of text data attract more and more scholars to study it. Text sentiment analysis has become a hot spot in social network data analysis and has important application value in academic field, social field, and business field. Based on the idea of pre-training, this paper improves the random word masking algorithm of deep pretraining task in the BERT (Bidirectional Encoder Representation) model to improve the efficiency and stability of model pretraining. Second, a new pretraining task of original sentence judgment is designed to enable the model to measure the degree of sentence smoothness, so that the BERT model can better understand the semantics of context. By referring to the idea of attention mechanism, a deep learning framework with attention weight added into gated convolution is constructed and the special attention weight method is adopted to enhance semantic information. Second, gating convolution and attention mechanism are combined to model aspect-related semantic information and text complete semantic information. Finally, classify the emotion classifier layer of social network, use Softmax function to complete negative, positive, and neutral multiple classifications and calculate the result of emotion classification. By applying the optimized convolutional neural network cyclic optimization network to single task and multitask in practice, the feasibility of applying the optimized convolutional neural network and cyclic neural network to social network sentiment analysis is verified.