摘要
公众社交媒体中,健康话题短文本存在特征维度稀疏、语义模糊、数据规模大等特点,导致其文本特征难以提取。对此,提出一种融合ERNIE和Bi-LSTM的融合模型ERNIE-Bi-LSTM,通过渐进式学习方法和双向注意力机制,提升健康话题短文本的分类效果。以微博、知乎、今日头条等7个社交媒体的热榜数据为实验对象,使用ERNIE模型完成预训练,利用BiLSTM双向注意力机制提取短文本词向量的特征,最终将获取的特征向量进行融合,并通过全连接层和Softmax分类器,获得短文本分类结果。实验结果表明,在真实社交媒体健康话题数据中,ERNIE-Bi-LSTM较ERNIE、Bert等4种文本分类模型具有较好的分类准确性,有效解决了健康话题短文本的分类问题。
In public social media,short texts on health topics are characterized by sparse feature dimension,fuzzy semantic meaning and large data scale,which makes it difficult to extract text features.In this paper,a fusion model ERNIE Bi-LSTM,which combines ERNIE and Bi-LSTM,is proposed to improve the classification effect of short texts on health topics through progressive learning method and two-way attention mechanism.This paper takes the hot list data of 7 social media such as Weibo,Zhihu and Toutiao as experimental objects,uses ERNIE model to complete pre-training,uses Bi-LSTM bidirectional attention mechanism to extract features of short text word vectors,and finally fuses the acquired feature vectors,and obtains short text classification results through the full connection layer and Softmax classifier.The experimental results show that in the real social media health topic data,the ERNIE-Bi-LSTM model has better classification accuracy than the four text classification models such as ERNIE and Bert,and effectively solves the classification problem of short texts on health topics.
作者
侯震
童惟依
邓靖飞
李扬
HOU Zhen;TONG Weiyi;DENG Jingfei;LI Yang(Institute of Medical Information,Chinese Academy of Medical Sciences,Beijing 100020,China)
出处
《电视技术》
2023年第7期18-23,27,共7页
Video Engineering
基金
2021年中国医学科学院院校创新工程项目(2021-I2M-1-033)。
