基于预训练语言模型的健康谣言检测  

Health Rumor Detection based on Pre-Trained Language Model

作　　者：许诺 赵薇[1] 尚柯源 陈浩宇 XU Nuo;ZHAO Wei;SHANG Keyuan;CHEN Haoyu(Communication University of China,Beijing 100024)

机构地区：[1]中国传媒大学,北京100024

出　　处：《系统科学与数学》2022年第10期2582-2589,共8页Journal of Systems Science and Mathematical Sciences

基　　金：中国传媒大学中央高校基本科研业务费专项(CUC220C008,CUC220B013)资助课题。

摘　　要：当前大多数谣言检测主要面向社交媒体数据,所处理文本序列较短,然而面向包含多个句子的段落或长序列文本篇章输入时,因不能提取有效特征进而影响模型识别效果.为获取谣言检测的有效信息,文章提出基于I-BERT-BiLSTM (Improved-BERT-BiLSTM)的健康类谣言检测方法,通过提取文档级长序列文本的摘要,并输入到以多层注意力机制为框架的深层神经网络进行特征提取,最后输入到BiLSTM进行谣言分类.实验结果表明:文章提出的I-BERT-BiLSTM模型在自建健康类谣言数据集与公开数据集上达到了97.75%和91.15%的准确率.Currently,most studies on rumor detection mainly focus on social media data and the length of text sequence is short.We argue that existing methods could not capture effective features from health rumors with long texts and then affect the validity of methods.To solve this,we propose an improved BERT-BiLSTM model(I-BERT-BiLSTM),which leverages effective information extracted from texts with long sequences for the health rumor detection.We first conduct text summarization from document-level text.The results are regarded as the input of the deep network model with multi-layer self-attention mechanisms for feature extraction.Finally,we feed the output into BiLSTM for rumor classification.The experimental results show that the model we proposed in this paper achieves 97.75% and 91.15% accuracy on the self-built health rumor data and public data.

关 键 词：谣言检测 预训练语言模型 摘要提取 I-BERT-BiLSTM 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]