2 Related Work
At present, there are two common methods for detecting microblog rumors, one is
feature classification based on traditional learning, and the other is rumor recognition
based on deep learning by constructing neural network model.
The feature classification method is mainly based on content features and user
features. Qazvinian et al. [1] used Bayes classifier to construct positive and negative
models and extracted various potential features such as behavioral features and emotional features of the information posted or put forwarded by users. Zhang et al. [2]
believe that shallow features cannot distinguish between rumor information and nonverbal information. Therefore, an automatic rumor detection method based on the
newly proposed implicit feature and the shallow feature of the message is proposed.
Compared with the common feature extraction classification algorithm, the neural
network model based on deep learning is more efficient and accurate for rumor
recognition. Ruchansky [3] and others proposed a model called CSI, which is based on
the improvement of recurrent neural network; the model is mainly divided into two
modules, the former captures the time response characteristics of users and content and
combines with the text, and the latter captured source features and achieved good
results in false news recognition.
The above research work had carried on rumor recognition from multiple dimensions and achieved good rumor recognition effect. However, the performance of
monolayer neural network model in feature extraction is not very ideal, and the
extraction of some deep features cannot be completed. When the common CNN is
combined with the LSTM network model, some user-content relationship characteristics are often neglected, which leads to the loss of some features, and the long-term
time series rules need to be optimized. Therefore, this paper proposes a CNN with
attention mechanism that combines with the LSTM network model to achieve the
purpose of rumor recognition.
3 Rumor Recognition Based on Composite Model
of User-Content Relations
3.1 Traditional CNN-LSTM
The convolutional neural network model based on deep learning has the ability of
learning characteristics hierarchically. Research [4] shows that the features learned by
convolutional neural networks have stronger discriminative ability and generalization
ability than artificial design features. The LSTM is a time recurrent neural network [5].
The neural network has three main gates (input gate, forgetting gate, and output gate)
and a cell unit component, which can realize time series prediction.
Traditional rumor detection method based on deep learning mainly extracts the
local features of short texts through the convolutional layer of CNN and constructs new
vectors by convolution kernel operation. On the characteristics of time dimension,
using LSTM network model with the ability of feature extraction for long-distance
432
Z. Zhou et al.
At present, there are two common methods for detecting microblog rumors, one is
feature classification based on traditional learning, and the other is rumor recognition
based on deep learning by constructing neural network model.
The feature classification method is mainly based on content features and user
features. Qazvinian et al. [1] used Bayes classifier to construct positive and negative
models and extracted various potential features such as behavioral features and emotional features of the information posted or put forwarded by users. Zhang et al. [2]
believe that shallow features cannot distinguish between rumor information and nonverbal information. Therefore, an automatic rumor detection method based on the
newly proposed implicit feature and the shallow feature of the message is proposed.
Compared with the common feature extraction classification algorithm, the neural
network model based on deep learning is more efficient and accurate for rumor
recognition. Ruchansky [3] and others proposed a model called CSI, which is based on
the improvement of recurrent neural network; the model is mainly divided into two
modules, the former captures the time response characteristics of users and content and
combines with the text, and the latter captured source features and achieved good
results in false news recognition.
The above research work had carried on rumor recognition from multiple dimensions and achieved good rumor recognition effect. However, the performance of
monolayer neural network model in feature extraction is not very ideal, and the
extraction of some deep features cannot be completed. When the common CNN is
combined with the LSTM network model, some user-content relationship characteristics are often neglected, which leads to the loss of some features, and the long-term
time series rules need to be optimized. Therefore, this paper proposes a CNN with
attention mechanism that combines with the LSTM network model to achieve the
purpose of rumor recognition.
3 Rumor Recognition Based on Composite Model
of User-Content Relations
3.1 Traditional CNN-LSTM
The convolutional neural network model based on deep learning has the ability of
learning characteristics hierarchically. Research [4] shows that the features learned by
convolutional neural networks have stronger discriminative ability and generalization
ability than artificial design features. The LSTM is a time recurrent neural network [5].
The neural network has three main gates (input gate, forgetting gate, and output gate)
and a cell unit component, which can realize time series prediction.
Traditional rumor detection method based on deep learning mainly extracts the
local features of short texts through the convolutional layer of CNN and constructs new
vectors by convolution kernel operation. On the characteristics of time dimension,
using LSTM network model with the ability of feature extraction for long-distance
432
Z. Zhou et al.