College students'sentiment analysis model combined with tag relevance
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
维普
万方数据
针对现有高校学生社交媒体评论情感分析易忽视标签相关性,以及多使用单一粒度特征学习文本表示的问题,提出一种融合标签相关性的高校学生情感分析模型.首先,利用双向Transformer编码器(Bidirectional Encoder Representations from Transformers,BERT)获取词向量表示,通过池化和双向长短期记忆网络(Bidirectional Long-Short Term Memory,Bi-LSTM)分别提取句子级与单词级的文本表示.然后,基于共现关系学习标签之间的相关性,通过句子级、单词级的"文本-标签"注意力获取特定于标签的特征表示,将特征进行融合,使用sigmoid分类器计算文本属于每一类情感标签的概率.实验结果表明,所提模型与对比模型相比,在汉明损失、排序损失和标签排序平均精度方面均有提高,验证了融合不同粒度文本特征与标签相关性对高校学生情感分析的有效性.
For the problem that existing sentiment analysis for social media comments by college students tend to ignore the tag relevance and mainly use single-granularity features to learn text representations,a multi-label text sentiment analysis model combined tag relevance is proposed.Bi-directional encoder representations from transformers(BERT)are used to create the word vector representations,and the pooling and bidirectional long-short term memory networks(Bi-LSTM)are used to extract the sentence-level and word-level text representations respectively.Correlations between labels are learned based on the co-occurrence of the labels.The feature representation for a certain label is obtained through the"text-label"attention process at the sentence and word levels,and the features are combined.A sigmoid classifier is used to determine the likelihood that a text belongs to each class of sentiment labels.Experiment results show that compared to the contrast models,the proposed model improves the Hamming loss,the ranking loss,and the label rank aver-age precision,which validates the effectiveness of integrating different granularity text features with tag relevance for college students'sentiment analysis.
维普
万方数据
