基于BERT模型的增强混合神经网络的谣言检测

网络社交平台中大量谣言的广泛传播严重影响社会稳定。传统谣言检测方法无法有效处理文本中多义词和突出重要关键词,造成检测效果不理想。针对该问题,提出一种基于BERT模型的增强混合神经网络的谣言检测方法。该方法使用BERT模型将推文向量化,通过3种不同尺寸的卷积核学习推文特征,将这些特征进行最大池化拼接得到特征序列,并输入到BiLSTM中学习序列特征。同时引入Attention机制计算注意力分值,实现谣言检测。在公开数据集Twitter15和Twitter16上的实验结果表明,该方法相较于其他方法在谣言检测性能方面提升明显,并在早期检测阶段表现出卓越的检测能力,检测准确率提高了5个百分点。

RUMOR DETECTION OF IMPROVED HYBIRD NEURAL NETWORK BASED ON BERT MODEL

There are a large number of rumors existing in social networking platforms,while the widespread of rumors have affected social stability seriously.The traditional rumor detection approaches cannot solve the problem of polysemy in the text and highlight the significant keywords,which leads to unsatisfactory detection results.To solve the problems,a rumor detection method based on BERT model and improved hybrid neural network is proposed.In this method,BERT model was used to vectorize the tweets.The convolution kernel of three different sizes was used to learn the features of tweets,and these features were added to the max pooling layer and spliced into feature sequences,which were inputted into BiLSTM for learning sequence features.Moreover,the Attention mechanism was added to calculate attention value.The results on Twitter15 and Twitter16 datasets demonstrate that the proposed method achieves much better performance than state-of-the-art approaches,and demonstrates superior capacity on detecting rumors at very early stage.The accuracy increases by 5 percent.